diff --git "a/parse/train/SyzVb3CcFX/SyzVb3CcFX_middle.json" "b/parse/train/SyzVb3CcFX/SyzVb3CcFX_middle.json"
new file mode 100644--- /dev/null
+++ "b/parse/train/SyzVb3CcFX/SyzVb3CcFX_middle.json"
@@ -0,0 +1,49159 @@
+{
+ "pdf_info": [
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 79,
+ 429,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 77,
+ 335,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 77,
+ 335,
+ 97
+ ],
+ "score": 1.0,
+ "content": "TIME-AGNOSTIC PREDICTION:",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 97,
+ 429,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 97,
+ 429,
+ 117
+ ],
+ "score": 1.0,
+ "content": "PREDICTING PREDICTABLE VIDEO FRAMES",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 114,
+ 135,
+ 196,
+ 158
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 114,
+ 134,
+ 198,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 134,
+ 198,
+ 148
+ ],
+ "score": 1.0,
+ "content": "Dinesh Jayaraman",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 112,
+ 145,
+ 168,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 145,
+ 168,
+ 159
+ ],
+ "score": 1.0,
+ "content": "UC Berkeley",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 227,
+ 135,
+ 293,
+ 157
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 226,
+ 134,
+ 294,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 226,
+ 134,
+ 294,
+ 147
+ ],
+ "score": 1.0,
+ "content": "Frederik Ebert",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 226,
+ 144,
+ 282,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 226,
+ 144,
+ 282,
+ 159
+ ],
+ "score": 1.0,
+ "content": "UC Berkeley",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 324,
+ 135,
+ 385,
+ 158
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 323,
+ 134,
+ 387,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 323,
+ 134,
+ 387,
+ 147
+ ],
+ "score": 1.0,
+ "content": "Alyosha Efros",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 323,
+ 145,
+ 379,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 323,
+ 145,
+ 379,
+ 159
+ ],
+ "score": 1.0,
+ "content": "UC Berkeley",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 416,
+ 136,
+ 478,
+ 158
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 415,
+ 133,
+ 480,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 415,
+ 133,
+ 480,
+ 149
+ ],
+ "score": 1.0,
+ "content": "Sergey Levine",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 415,
+ 145,
+ 471,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 415,
+ 145,
+ 471,
+ 159
+ ],
+ "score": 1.0,
+ "content": "UC Berkeley",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 187,
+ 333,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 185,
+ 336,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 185,
+ 336,
+ 201
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 209,
+ 469,
+ 351
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 208,
+ 469,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 208,
+ 469,
+ 222
+ ],
+ "score": 1.0,
+ "content": "Prediction is arguably one of the most basic functions of an intelligent system. In",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 140,
+ 219,
+ 470,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 140,
+ 219,
+ 470,
+ 234
+ ],
+ "score": 1.0,
+ "content": "general, the problem of predicting events in the future or between two waypoints is",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 231,
+ 470,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 231,
+ 470,
+ 244
+ ],
+ "score": 1.0,
+ "content": "exceedingly difficult. However, most phenomena naturally pass through relatively",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 242,
+ 470,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 242,
+ 470,
+ 254
+ ],
+ "score": 1.0,
+ "content": "predictable bottlenecks—while we cannot predict the precise trajectory of a robot",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 253,
+ 470,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 253,
+ 470,
+ 265
+ ],
+ "score": 1.0,
+ "content": "arm between being at rest and holding an object up, we can be certain that it",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 141,
+ 264,
+ 469,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 264,
+ 469,
+ 276
+ ],
+ "score": 1.0,
+ "content": "must have picked the object up. To exploit this, we decouple visual prediction",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 275,
+ 470,
+ 287
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 275,
+ 470,
+ 287
+ ],
+ "score": 1.0,
+ "content": "from a rigid notion of time. While conventional approaches predict frames at",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 141,
+ 286,
+ 470,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 286,
+ 470,
+ 299
+ ],
+ "score": 1.0,
+ "content": "regularly spaced temporal intervals, our time-agnostic predictors (TAP) are not tied",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 298,
+ 470,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 298,
+ 470,
+ 308
+ ],
+ "score": 1.0,
+ "content": "to specific times so that they may instead discover predictable “bottleneck” frames",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 308,
+ 469,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 308,
+ 469,
+ 320
+ ],
+ "score": 1.0,
+ "content": "no matter when they occur. We evaluate our approach for future and intermediate",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 142,
+ 319,
+ 469,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 319,
+ 469,
+ 331
+ ],
+ "score": 1.0,
+ "content": "frame prediction across three robotic manipulation tasks. Our predictions are not",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 330,
+ 470,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 330,
+ 470,
+ 343
+ ],
+ "score": 1.0,
+ "content": "only of higher visual quality, but also correspond to coherent semantic subgoals in",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 141,
+ 341,
+ 252,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 341,
+ 252,
+ 353
+ ],
+ "score": 1.0,
+ "content": "temporally extended tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 367,
+ 206,
+ 379
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 208,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 208,
+ 383
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 389,
+ 505,
+ 434
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 390,
+ 506,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 390,
+ 506,
+ 402
+ ],
+ "score": 1.0,
+ "content": "Imagine taking a bottle of water and laying it on its side. Consider what happens to the surface of",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 401,
+ 505,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 401,
+ 505,
+ 413
+ ],
+ "score": 1.0,
+ "content": "the water as you do this: which times can you confidently make predictions about? The surface is",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 411,
+ 506,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 411,
+ 506,
+ 424
+ ],
+ "score": 1.0,
+ "content": "initially flat, then becomes turbulent, until it is flat again, as shown in Fig 1. Predicting the exact",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 423,
+ 506,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 423,
+ 506,
+ 435
+ ],
+ "score": 1.0,
+ "content": "shape of the turbulent liquid is extremely hard, but its easy to say that it will eventually settle down.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 26.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 439,
+ 505,
+ 538
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 440,
+ 505,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 440,
+ 505,
+ 452
+ ],
+ "score": 1.0,
+ "content": "Prediction is thought to be fundamental to intelligence (Bar, 2009; Clark, 2013; Hohwy, 2013). If an",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 450,
+ 506,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 506,
+ 463
+ ],
+ "score": 1.0,
+ "content": "agent can learn to predict the future, it can take anticipatory actions, plan through its predictions, and",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 461,
+ 506,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 506,
+ 475
+ ],
+ "score": 1.0,
+ "content": "use prediction as a proxy for representation learning. The key difficulty in prediction is uncertainty.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "Visual prediction approaches attempt to mitigate uncertainty by predicting iteratively in heuristically",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 483,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 483,
+ 253,
+ 497
+ ],
+ "score": 1.0,
+ "content": "chosen small timesteps, such as, say,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 254,
+ 484,
+ 272,
+ 494
+ ],
+ "score": 0.41,
+ "content": "0 . 1 s",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 273,
+ 483,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": ". In the bottle-tilting case, such approaches generate blurry",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 494,
+ 506,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 226,
+ 507
+ ],
+ "score": 1.0,
+ "content": "images of the chaotic states at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 226,
+ 495,
+ 300,
+ 506
+ ],
+ "score": 0.75,
+ "content": "t = 0 . 1 s , 0 . 2 s , . . . ,",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 300,
+ 494,
+ 506,
+ 507
+ ],
+ "score": 1.0,
+ "content": ", and this blurriness compounds to make predictions",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 505,
+ 506,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 506,
+ 518
+ ],
+ "score": 1.0,
+ "content": "unusable within a few steps. Sophisticated probabilistic approaches have been proposed to better",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 516,
+ 506,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 516,
+ 506,
+ 529
+ ],
+ "score": 1.0,
+ "content": "handle this uncertainty (Babaeizadeh et al., 2018; Lee et al., 2018; Denton & Fergus, 2018; Xue et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 526,
+ 136,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 526,
+ 136,
+ 541
+ ],
+ "score": 1.0,
+ "content": "2016).",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 544,
+ 505,
+ 621
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 544,
+ 505,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 544,
+ 505,
+ 557
+ ],
+ "score": 1.0,
+ "content": "What if we instead change the goal of our prediction models? Fixed time intervals in prediction are",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 556,
+ 505,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 556,
+ 505,
+ 568
+ ],
+ "score": 1.0,
+ "content": "in many ways an artifact of the fact that cameras and monitors record and display video at fixed",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 566,
+ 505,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 505,
+ 578
+ ],
+ "score": 1.0,
+ "content": "frequencies. Rather than requiring predictions at regularly spaced future frames, we ask: if a frame",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 590
+ ],
+ "score": 1.0,
+ "content": "prediction is treated as a bet on that frame occurring at some future point, what should we predict?",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 601
+ ],
+ "score": 1.0,
+ "content": "Such time-agnostic prediction (TAP) has two immediate effects: (i) the predictor may skip more",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 600,
+ 504,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 504,
+ 610
+ ],
+ "score": 1.0,
+ "content": "uncertain states in favor of less uncertain ones, and (ii) while in the standard approach, a prediction is",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 610,
+ 506,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 610,
+ 187,
+ 623
+ ],
+ "score": 1.0,
+ "content": "wrong if it occurs at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 188,
+ 611,
+ 209,
+ 620
+ ],
+ "score": 0.87,
+ "content": "t \\pm \\epsilon",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 209,
+ 610,
+ 264,
+ 623
+ ],
+ "score": 1.0,
+ "content": "rather than at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 611,
+ 269,
+ 620
+ ],
+ "score": 0.73,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 270,
+ 610,
+ 506,
+ 623
+ ],
+ "score": 1.0,
+ "content": ", our formulation considers such predictions equally correct.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 41
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 627,
+ 505,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 639
+ ],
+ "score": 1.0,
+ "content": "Recall the bottle-tilting uncertainty profile. Fig 1 depicts uncertainty profiles for several other",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "prediction settings, including both forward/future prediction (given a start frame) and intermediate",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 648,
+ 506,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 506,
+ 662
+ ],
+ "score": 1.0,
+ "content": "prediction (given start and end frames). Our time-agnostic reframing of the prediction problem targets",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 659,
+ 505,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 659,
+ 505,
+ 673
+ ],
+ "score": 1.0,
+ "content": "the minima of these profiles, where prediction is intuitively easiest. We refer to these minima states",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 175,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 175,
+ 682
+ ],
+ "score": 1.0,
+ "content": "as “bottlenecks.”",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 47
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "At this point, one might ask: are these “easy” bottlenecks actually useful to predict? Intuitively,",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "bottlenecks naturally correspond to reliable subgoals—an agent hoping to solve the maze in Fig 1 (e)",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "would do well to target its bottlenecks as subgoals. In our experiments, we evaluate the usefulness of",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 379,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 379,
+ 732
+ ],
+ "score": 1.0,
+ "content": "our predictions as subgoals in simulated robotic manipulation tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ }
+ ],
+ "index": 51.5
+ }
+ ],
+ "page_idx": 0,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "1",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 79,
+ 429,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 77,
+ 335,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 77,
+ 335,
+ 97
+ ],
+ "score": 1.0,
+ "content": "TIME-AGNOSTIC PREDICTION:",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 97,
+ 429,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 97,
+ 429,
+ 117
+ ],
+ "score": 1.0,
+ "content": "PREDICTING PREDICTABLE VIDEO FRAMES",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 114,
+ 135,
+ 196,
+ 158
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 114,
+ 134,
+ 198,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 134,
+ 198,
+ 148
+ ],
+ "score": 1.0,
+ "content": "Dinesh Jayaraman",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 112,
+ 145,
+ 168,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 145,
+ 168,
+ 159
+ ],
+ "score": 1.0,
+ "content": "UC Berkeley",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.5,
+ "bbox_fs": [
+ 112,
+ 134,
+ 198,
+ 159
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 227,
+ 135,
+ 293,
+ 157
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 226,
+ 134,
+ 294,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 226,
+ 134,
+ 294,
+ 147
+ ],
+ "score": 1.0,
+ "content": "Frederik Ebert",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 226,
+ 144,
+ 282,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 226,
+ 144,
+ 282,
+ 159
+ ],
+ "score": 1.0,
+ "content": "UC Berkeley",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4.5,
+ "bbox_fs": [
+ 226,
+ 134,
+ 294,
+ 159
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 324,
+ 135,
+ 385,
+ 158
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 323,
+ 134,
+ 387,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 323,
+ 134,
+ 387,
+ 147
+ ],
+ "score": 1.0,
+ "content": "Alyosha Efros",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 323,
+ 145,
+ 379,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 323,
+ 145,
+ 379,
+ 159
+ ],
+ "score": 1.0,
+ "content": "UC Berkeley",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6.5,
+ "bbox_fs": [
+ 323,
+ 134,
+ 387,
+ 159
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 416,
+ 136,
+ 478,
+ 158
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 415,
+ 133,
+ 480,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 415,
+ 133,
+ 480,
+ 149
+ ],
+ "score": 1.0,
+ "content": "Sergey Levine",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 415,
+ 145,
+ 471,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 415,
+ 145,
+ 471,
+ 159
+ ],
+ "score": 1.0,
+ "content": "UC Berkeley",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8.5,
+ "bbox_fs": [
+ 415,
+ 133,
+ 480,
+ 159
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 187,
+ 333,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 185,
+ 336,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 185,
+ 336,
+ 201
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 209,
+ 469,
+ 351
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 208,
+ 469,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 208,
+ 469,
+ 222
+ ],
+ "score": 1.0,
+ "content": "Prediction is arguably one of the most basic functions of an intelligent system. In",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 140,
+ 219,
+ 470,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 140,
+ 219,
+ 470,
+ 234
+ ],
+ "score": 1.0,
+ "content": "general, the problem of predicting events in the future or between two waypoints is",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 231,
+ 470,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 231,
+ 470,
+ 244
+ ],
+ "score": 1.0,
+ "content": "exceedingly difficult. However, most phenomena naturally pass through relatively",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 242,
+ 470,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 242,
+ 470,
+ 254
+ ],
+ "score": 1.0,
+ "content": "predictable bottlenecks—while we cannot predict the precise trajectory of a robot",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 253,
+ 470,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 253,
+ 470,
+ 265
+ ],
+ "score": 1.0,
+ "content": "arm between being at rest and holding an object up, we can be certain that it",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 141,
+ 264,
+ 469,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 264,
+ 469,
+ 276
+ ],
+ "score": 1.0,
+ "content": "must have picked the object up. To exploit this, we decouple visual prediction",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 275,
+ 470,
+ 287
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 275,
+ 470,
+ 287
+ ],
+ "score": 1.0,
+ "content": "from a rigid notion of time. While conventional approaches predict frames at",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 141,
+ 286,
+ 470,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 286,
+ 470,
+ 299
+ ],
+ "score": 1.0,
+ "content": "regularly spaced temporal intervals, our time-agnostic predictors (TAP) are not tied",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 298,
+ 470,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 298,
+ 470,
+ 308
+ ],
+ "score": 1.0,
+ "content": "to specific times so that they may instead discover predictable “bottleneck” frames",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 308,
+ 469,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 308,
+ 469,
+ 320
+ ],
+ "score": 1.0,
+ "content": "no matter when they occur. We evaluate our approach for future and intermediate",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 142,
+ 319,
+ 469,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 319,
+ 469,
+ 331
+ ],
+ "score": 1.0,
+ "content": "frame prediction across three robotic manipulation tasks. Our predictions are not",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 330,
+ 470,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 330,
+ 470,
+ 343
+ ],
+ "score": 1.0,
+ "content": "only of higher visual quality, but also correspond to coherent semantic subgoals in",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 141,
+ 341,
+ 252,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 341,
+ 252,
+ 353
+ ],
+ "score": 1.0,
+ "content": "temporally extended tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 17,
+ "bbox_fs": [
+ 140,
+ 208,
+ 470,
+ 353
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 367,
+ 206,
+ 379
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 208,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 208,
+ 383
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 389,
+ 505,
+ 434
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 390,
+ 506,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 390,
+ 506,
+ 402
+ ],
+ "score": 1.0,
+ "content": "Imagine taking a bottle of water and laying it on its side. Consider what happens to the surface of",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 401,
+ 505,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 401,
+ 505,
+ 413
+ ],
+ "score": 1.0,
+ "content": "the water as you do this: which times can you confidently make predictions about? The surface is",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 411,
+ 506,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 411,
+ 506,
+ 424
+ ],
+ "score": 1.0,
+ "content": "initially flat, then becomes turbulent, until it is flat again, as shown in Fig 1. Predicting the exact",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 423,
+ 506,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 423,
+ 506,
+ 435
+ ],
+ "score": 1.0,
+ "content": "shape of the turbulent liquid is extremely hard, but its easy to say that it will eventually settle down.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 105,
+ 390,
+ 506,
+ 435
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 439,
+ 505,
+ 538
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 440,
+ 505,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 440,
+ 505,
+ 452
+ ],
+ "score": 1.0,
+ "content": "Prediction is thought to be fundamental to intelligence (Bar, 2009; Clark, 2013; Hohwy, 2013). If an",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 450,
+ 506,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 506,
+ 463
+ ],
+ "score": 1.0,
+ "content": "agent can learn to predict the future, it can take anticipatory actions, plan through its predictions, and",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 461,
+ 506,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 506,
+ 475
+ ],
+ "score": 1.0,
+ "content": "use prediction as a proxy for representation learning. The key difficulty in prediction is uncertainty.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "Visual prediction approaches attempt to mitigate uncertainty by predicting iteratively in heuristically",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 483,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 483,
+ 253,
+ 497
+ ],
+ "score": 1.0,
+ "content": "chosen small timesteps, such as, say,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 254,
+ 484,
+ 272,
+ 494
+ ],
+ "score": 0.41,
+ "content": "0 . 1 s",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 273,
+ 483,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": ". In the bottle-tilting case, such approaches generate blurry",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 494,
+ 506,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 226,
+ 507
+ ],
+ "score": 1.0,
+ "content": "images of the chaotic states at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 226,
+ 495,
+ 300,
+ 506
+ ],
+ "score": 0.75,
+ "content": "t = 0 . 1 s , 0 . 2 s , . . . ,",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 300,
+ 494,
+ 506,
+ 507
+ ],
+ "score": 1.0,
+ "content": ", and this blurriness compounds to make predictions",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 505,
+ 506,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 506,
+ 518
+ ],
+ "score": 1.0,
+ "content": "unusable within a few steps. Sophisticated probabilistic approaches have been proposed to better",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 516,
+ 506,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 516,
+ 506,
+ 529
+ ],
+ "score": 1.0,
+ "content": "handle this uncertainty (Babaeizadeh et al., 2018; Lee et al., 2018; Denton & Fergus, 2018; Xue et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 526,
+ 136,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 526,
+ 136,
+ 541
+ ],
+ "score": 1.0,
+ "content": "2016).",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 33,
+ "bbox_fs": [
+ 105,
+ 440,
+ 506,
+ 541
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 544,
+ 505,
+ 621
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 544,
+ 505,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 544,
+ 505,
+ 557
+ ],
+ "score": 1.0,
+ "content": "What if we instead change the goal of our prediction models? Fixed time intervals in prediction are",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 556,
+ 505,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 556,
+ 505,
+ 568
+ ],
+ "score": 1.0,
+ "content": "in many ways an artifact of the fact that cameras and monitors record and display video at fixed",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 566,
+ 505,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 566,
+ 505,
+ 578
+ ],
+ "score": 1.0,
+ "content": "frequencies. Rather than requiring predictions at regularly spaced future frames, we ask: if a frame",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 590
+ ],
+ "score": 1.0,
+ "content": "prediction is treated as a bet on that frame occurring at some future point, what should we predict?",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 601
+ ],
+ "score": 1.0,
+ "content": "Such time-agnostic prediction (TAP) has two immediate effects: (i) the predictor may skip more",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 600,
+ 504,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 504,
+ 610
+ ],
+ "score": 1.0,
+ "content": "uncertain states in favor of less uncertain ones, and (ii) while in the standard approach, a prediction is",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 610,
+ 506,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 610,
+ 187,
+ 623
+ ],
+ "score": 1.0,
+ "content": "wrong if it occurs at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 188,
+ 611,
+ 209,
+ 620
+ ],
+ "score": 0.87,
+ "content": "t \\pm \\epsilon",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 209,
+ 610,
+ 264,
+ 623
+ ],
+ "score": 1.0,
+ "content": "rather than at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 264,
+ 611,
+ 269,
+ 620
+ ],
+ "score": 0.73,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 270,
+ 610,
+ 506,
+ 623
+ ],
+ "score": 1.0,
+ "content": ", our formulation considers such predictions equally correct.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 41,
+ "bbox_fs": [
+ 105,
+ 544,
+ 506,
+ 623
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 627,
+ 505,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 639
+ ],
+ "score": 1.0,
+ "content": "Recall the bottle-tilting uncertainty profile. Fig 1 depicts uncertainty profiles for several other",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "prediction settings, including both forward/future prediction (given a start frame) and intermediate",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 648,
+ 506,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 506,
+ 662
+ ],
+ "score": 1.0,
+ "content": "prediction (given start and end frames). Our time-agnostic reframing of the prediction problem targets",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 659,
+ 505,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 659,
+ 505,
+ 673
+ ],
+ "score": 1.0,
+ "content": "the minima of these profiles, where prediction is intuitively easiest. We refer to these minima states",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 175,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 175,
+ 682
+ ],
+ "score": 1.0,
+ "content": "as “bottlenecks.”",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 47,
+ "bbox_fs": [
+ 105,
+ 627,
+ 506,
+ 682
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "At this point, one might ask: are these “easy” bottlenecks actually useful to predict? Intuitively,",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "bottlenecks naturally correspond to reliable subgoals—an agent hoping to solve the maze in Fig 1 (e)",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "would do well to target its bottlenecks as subgoals. In our experiments, we evaluate the usefulness of",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 379,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 379,
+ 732
+ ],
+ "score": 1.0,
+ "content": "our predictions as subgoals in simulated robotic manipulation tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ }
+ ],
+ "index": 51.5,
+ "bbox_fs": [
+ 105,
+ 687,
+ 506,
+ 732
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 106,
+ 75,
+ 503,
+ 155
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 106,
+ 75,
+ 503,
+ 155
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 75,
+ 503,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 75,
+ 503,
+ 155
+ ],
+ "score": 0.966,
+ "type": "image",
+ "image_path": "92f28243dd5787bef0b1258b64d2a01ce30a0854c5c61a880096ebae7320225f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 75,
+ 503,
+ 101.66666666666667
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 101.66666666666667,
+ 503,
+ 128.33333333333334
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 128.33333333333334,
+ 503,
+ 155.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 163,
+ 505,
+ 213
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 163,
+ 504,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 163,
+ 504,
+ 173
+ ],
+ "score": 1.0,
+ "content": "Figure 1: (a) Over time as the bottle is tilted, the uncertainty first rises and then falls as the bottle is held steady",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 172,
+ 506,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 172,
+ 506,
+ 184
+ ],
+ "score": 1.0,
+ "content": "after tilting. (b)-(e) Similar uncertainty profiles corresponding to various scenarios—a ball rolling down the side",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 181,
+ 506,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 181,
+ 286,
+ 195
+ ],
+ "score": 1.0,
+ "content": "of a bowl, a car driving on a highway with an exit",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 286,
+ 183,
+ 307,
+ 192
+ ],
+ "score": 0.63,
+ "content": "1 0 0 \\mathrm { m }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 307,
+ 181,
+ 506,
+ 195
+ ],
+ "score": 1.0,
+ "content": "away, an iron pellet tossed in the direction of a magnet,",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 192,
+ 505,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 192,
+ 505,
+ 205
+ ],
+ "score": 1.0,
+ "content": "and intermediate frame prediction in a maze traversal given start and end states. The red asterisks along the",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 202,
+ 505,
+ 214
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 204,
+ 112,
+ 212
+ ],
+ "score": 0.47,
+ "content": "\\mathbf { X } ^ { } -",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 113,
+ 202,
+ 505,
+ 214
+ ],
+ "score": 1.0,
+ "content": "-axis correspond to the asterisks in the maze—these “bottleneck” states must occur in any successful traversal.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 225,
+ 505,
+ 269
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 224,
+ 506,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 224,
+ 506,
+ 237
+ ],
+ "score": 1.0,
+ "content": "Our main contributions are: (i) we reframe the video prediction problem to be time-agnostic, (ii) we",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 236,
+ 506,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 236,
+ 506,
+ 249
+ ],
+ "score": 1.0,
+ "content": "propose a novel technical approach to solve this problem, (iii) we show that our approach effectively",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 246,
+ 505,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 246,
+ 505,
+ 259
+ ],
+ "score": 1.0,
+ "content": "identifies “bottleneck states” across several tasks, and (iv) we show that these bottlenecks correspond",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 258,
+ 347,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 347,
+ 271
+ ],
+ "score": 1.0,
+ "content": "to subgoals that aid in planning towards complex end goals.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 286,
+ 211,
+ 299
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 285,
+ 213,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 285,
+ 213,
+ 302
+ ],
+ "score": 1.0,
+ "content": "2 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 308,
+ 506,
+ 462
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 506,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 506,
+ 321
+ ],
+ "score": 1.0,
+ "content": "Visual prediction approaches. Prior visual prediction approaches regress directly to future video",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 321,
+ 505,
+ 332
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 321,
+ 505,
+ 332
+ ],
+ "score": 1.0,
+ "content": "frames in the pixel space (Ranzato et al., 2014; Oh et al., 2015) or in a learned feature space (Hadsell",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 331,
+ 506,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 331,
+ 506,
+ 343
+ ],
+ "score": 1.0,
+ "content": "et al., 2006; Mobahi et al., 2009; Jayaraman & Grauman, 2015; Wang et al., 2016; Vondrick et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 341,
+ 506,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 341,
+ 506,
+ 355
+ ],
+ "score": 1.0,
+ "content": "2016b; Kitani et al., 2012). The success of generative adversarial networks (GANs) (Goodfellow",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 104,
+ 351,
+ 506,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 351,
+ 506,
+ 367
+ ],
+ "score": 1.0,
+ "content": "et al., 2014; Mirza & Osindero, 2014; Radford et al., 2015; Isola et al., 2017) has inspired many",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 364,
+ 505,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 364,
+ 505,
+ 375
+ ],
+ "score": 1.0,
+ "content": "video prediction approaches (Mathieu et al., 2015; Vondrick et al., 2016a; Finn & Levine, 2017; Xue",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 375,
+ 507,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 375,
+ 507,
+ 387
+ ],
+ "score": 1.0,
+ "content": "et al., 2016; Oh et al., 2015; Ebert et al., 2017; Finn et al., 2016; Larsen et al., 2016; Lee et al., 2018).",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 386,
+ 506,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 386,
+ 506,
+ 399
+ ],
+ "score": 1.0,
+ "content": "While adversarial losses aid in producing photorealistic image patches, prediction has to contend",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 506,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 506,
+ 409
+ ],
+ "score": 1.0,
+ "content": "with a more fundamental problem: uncertainty. Several approaches (Walker et al., 2016; Xue et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 407,
+ 506,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 407,
+ 506,
+ 420
+ ],
+ "score": 1.0,
+ "content": "2016; Denton & Fergus, 2018; Lee et al., 2018; Larsen et al., 2016; Babaeizadeh et al., 2018) exploit",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 418,
+ 506,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 418,
+ 506,
+ 431
+ ],
+ "score": 1.0,
+ "content": "conditional variational autoencoders (VAE) (Kingma & Welling, 2013) to train latent variable models",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 429,
+ 506,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 429,
+ 506,
+ 442
+ ],
+ "score": 1.0,
+ "content": "for video prediction. Pixel-autoregression (Oord et al., 2016; van den Oord et al., 2016; Kalchbrenner",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 441,
+ 505,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 441,
+ 505,
+ 453
+ ],
+ "score": 1.0,
+ "content": "et al., 2016) explicitly factorizes the joint distribution over all pixels to model uncertainty, at a high",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 452,
+ 187,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 452,
+ 187,
+ 464
+ ],
+ "score": 1.0,
+ "content": "computational cost.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 19.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 468,
+ 505,
+ 524
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 481
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 481
+ ],
+ "score": 1.0,
+ "content": "Like these prior approaches, we too address the uncertainty problem in video prediction. We propose",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 479,
+ 506,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 479,
+ 506,
+ 492
+ ],
+ "score": 1.0,
+ "content": "a general time-agnostic prediction (TAP) framework for prediction tasks. While all prior work",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 490,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 490,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "predicts at fixed time intervals, we aim to identify inherently low-uncertainty bottleneck frames with",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 502,
+ 505,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 502,
+ 505,
+ 514
+ ],
+ "score": 1.0,
+ "content": "no associated timestamp. We show how TAP may be combined with conditional GANs as well as",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 512,
+ 340,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 512,
+ 340,
+ 525
+ ],
+ "score": 1.0,
+ "content": "VAEs, to handle the residual uncertainty in its predictions.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 541,
+ 505,
+ 650
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 506,
+ 553
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 506,
+ 553
+ ],
+ "score": 1.0,
+ "content": "Bottlenecks. In hierarchical reinforcement learning, bottlenecks are proposed for discovery of",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 552,
+ 506,
+ 564
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 552,
+ 506,
+ 564
+ ],
+ "score": 1.0,
+ "content": "options (Sutton et al., 1999) in low-dimensional state spaces in (McGovern & Barto, 2001; S¸ ims¸ek",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 562,
+ 506,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 562,
+ 506,
+ 576
+ ],
+ "score": 1.0,
+ "content": "& Barto, 2009; Bacon, 2013; Metzen, 2013). Most approaches (S¸ims¸ek & Barto, 2009; Bacon,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 505,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 505,
+ 586
+ ],
+ "score": 1.0,
+ "content": "2013; Metzen, 2013) construct full transition graphs and apply notions of graph centrality to locate",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 584,
+ 505,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 584,
+ 505,
+ 598
+ ],
+ "score": 1.0,
+ "content": "bottlenecks. A multi-instance learning approach is applied in (McGovern & Barto, 2001) to mine",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 595,
+ 505,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 595,
+ 505,
+ 607
+ ],
+ "score": 1.0,
+ "content": "states that occur in successful trajectories but not in others. We consider the use of our bottleneck",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 104,
+ 606,
+ 506,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 606,
+ 506,
+ 619
+ ],
+ "score": 1.0,
+ "content": "predictions as subgoals for a hierarchical planner, which is loosely related to options in that both aim",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 630
+ ],
+ "score": 1.0,
+ "content": "to break down temporally extended trajectories into more manageable chunks. Unlike these prior",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 641
+ ],
+ "score": 1.0,
+ "content": "works, we use predictability to identify bottlenecks, and apply this to unlabeled high-dimensional",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 640,
+ 202,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 640,
+ 202,
+ 651
+ ],
+ "score": 1.0,
+ "content": "visual state trajectories.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 36.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 656,
+ 505,
+ 733
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 656,
+ 505,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 656,
+ 505,
+ 669
+ ],
+ "score": 1.0,
+ "content": "Concurrently with us, Neitz et al. (2018) also propose a similar idea that allows a predictor to select",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 667,
+ 505,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 667,
+ 505,
+ 680
+ ],
+ "score": 1.0,
+ "content": "when to predict, and their experiments demonstrate its advantages in specially constructed tasks with",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 678,
+ 506,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 678,
+ 506,
+ 691
+ ],
+ "score": 1.0,
+ "content": "clear bottlenecks. In comparison, we propose not just the basic time-agnostic loss (Sec 3.1), but also",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 689,
+ 505,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 689,
+ 505,
+ 702
+ ],
+ "score": 1.0,
+ "content": "improvements in Sec 3.2 through 3.5 that allow time-agnostic prediction to work in more general",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "tasks such as synthetic and real videos of robotic object manipulation. Our experiments also test the",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 711,
+ 506,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 711,
+ 506,
+ 724
+ ],
+ "score": 1.0,
+ "content": "quality of discovered bottlenecks in these scenarios and their usefulness as subgoals for hierarchical",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 104,
+ 721,
+ 147,
+ 737
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 721,
+ 147,
+ 737
+ ],
+ "score": 1.0,
+ "content": "planning.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "page_idx": 1,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 763
+ ],
+ "score": 1.0,
+ "content": "2",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 106,
+ 75,
+ 503,
+ 155
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 106,
+ 75,
+ 503,
+ 155
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 75,
+ 503,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 75,
+ 503,
+ 155
+ ],
+ "score": 0.966,
+ "type": "image",
+ "image_path": "92f28243dd5787bef0b1258b64d2a01ce30a0854c5c61a880096ebae7320225f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 75,
+ 503,
+ 101.66666666666667
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 101.66666666666667,
+ 503,
+ 128.33333333333334
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 128.33333333333334,
+ 503,
+ 155.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 163,
+ 505,
+ 213
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 163,
+ 504,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 163,
+ 504,
+ 173
+ ],
+ "score": 1.0,
+ "content": "Figure 1: (a) Over time as the bottle is tilted, the uncertainty first rises and then falls as the bottle is held steady",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 172,
+ 506,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 172,
+ 506,
+ 184
+ ],
+ "score": 1.0,
+ "content": "after tilting. (b)-(e) Similar uncertainty profiles corresponding to various scenarios—a ball rolling down the side",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 181,
+ 506,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 181,
+ 286,
+ 195
+ ],
+ "score": 1.0,
+ "content": "of a bowl, a car driving on a highway with an exit",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 286,
+ 183,
+ 307,
+ 192
+ ],
+ "score": 0.63,
+ "content": "1 0 0 \\mathrm { m }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 307,
+ 181,
+ 506,
+ 195
+ ],
+ "score": 1.0,
+ "content": "away, an iron pellet tossed in the direction of a magnet,",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 192,
+ 505,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 192,
+ 505,
+ 205
+ ],
+ "score": 1.0,
+ "content": "and intermediate frame prediction in a maze traversal given start and end states. The red asterisks along the",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 202,
+ 505,
+ 214
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 204,
+ 112,
+ 212
+ ],
+ "score": 0.47,
+ "content": "\\mathbf { X } ^ { } -",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 113,
+ 202,
+ 505,
+ 214
+ ],
+ "score": 1.0,
+ "content": "-axis correspond to the asterisks in the maze—these “bottleneck” states must occur in any successful traversal.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 225,
+ 505,
+ 269
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 224,
+ 506,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 224,
+ 506,
+ 237
+ ],
+ "score": 1.0,
+ "content": "Our main contributions are: (i) we reframe the video prediction problem to be time-agnostic, (ii) we",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 236,
+ 506,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 236,
+ 506,
+ 249
+ ],
+ "score": 1.0,
+ "content": "propose a novel technical approach to solve this problem, (iii) we show that our approach effectively",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 246,
+ 505,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 246,
+ 505,
+ 259
+ ],
+ "score": 1.0,
+ "content": "identifies “bottleneck states” across several tasks, and (iv) we show that these bottlenecks correspond",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 258,
+ 347,
+ 271
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 347,
+ 271
+ ],
+ "score": 1.0,
+ "content": "to subgoals that aid in planning towards complex end goals.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9.5,
+ "bbox_fs": [
+ 105,
+ 224,
+ 506,
+ 271
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 286,
+ 211,
+ 299
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 285,
+ 213,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 285,
+ 213,
+ 302
+ ],
+ "score": 1.0,
+ "content": "2 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 308,
+ 506,
+ 462
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 506,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 506,
+ 321
+ ],
+ "score": 1.0,
+ "content": "Visual prediction approaches. Prior visual prediction approaches regress directly to future video",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 321,
+ 505,
+ 332
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 321,
+ 505,
+ 332
+ ],
+ "score": 1.0,
+ "content": "frames in the pixel space (Ranzato et al., 2014; Oh et al., 2015) or in a learned feature space (Hadsell",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 331,
+ 506,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 331,
+ 506,
+ 343
+ ],
+ "score": 1.0,
+ "content": "et al., 2006; Mobahi et al., 2009; Jayaraman & Grauman, 2015; Wang et al., 2016; Vondrick et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 341,
+ 506,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 341,
+ 506,
+ 355
+ ],
+ "score": 1.0,
+ "content": "2016b; Kitani et al., 2012). The success of generative adversarial networks (GANs) (Goodfellow",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 104,
+ 351,
+ 506,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 351,
+ 506,
+ 367
+ ],
+ "score": 1.0,
+ "content": "et al., 2014; Mirza & Osindero, 2014; Radford et al., 2015; Isola et al., 2017) has inspired many",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 364,
+ 505,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 364,
+ 505,
+ 375
+ ],
+ "score": 1.0,
+ "content": "video prediction approaches (Mathieu et al., 2015; Vondrick et al., 2016a; Finn & Levine, 2017; Xue",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 375,
+ 507,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 375,
+ 507,
+ 387
+ ],
+ "score": 1.0,
+ "content": "et al., 2016; Oh et al., 2015; Ebert et al., 2017; Finn et al., 2016; Larsen et al., 2016; Lee et al., 2018).",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 386,
+ 506,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 386,
+ 506,
+ 399
+ ],
+ "score": 1.0,
+ "content": "While adversarial losses aid in producing photorealistic image patches, prediction has to contend",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 506,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 506,
+ 409
+ ],
+ "score": 1.0,
+ "content": "with a more fundamental problem: uncertainty. Several approaches (Walker et al., 2016; Xue et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 407,
+ 506,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 407,
+ 506,
+ 420
+ ],
+ "score": 1.0,
+ "content": "2016; Denton & Fergus, 2018; Lee et al., 2018; Larsen et al., 2016; Babaeizadeh et al., 2018) exploit",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 418,
+ 506,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 418,
+ 506,
+ 431
+ ],
+ "score": 1.0,
+ "content": "conditional variational autoencoders (VAE) (Kingma & Welling, 2013) to train latent variable models",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 429,
+ 506,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 429,
+ 506,
+ 442
+ ],
+ "score": 1.0,
+ "content": "for video prediction. Pixel-autoregression (Oord et al., 2016; van den Oord et al., 2016; Kalchbrenner",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 441,
+ 505,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 441,
+ 505,
+ 453
+ ],
+ "score": 1.0,
+ "content": "et al., 2016) explicitly factorizes the joint distribution over all pixels to model uncertainty, at a high",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 452,
+ 187,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 452,
+ 187,
+ 464
+ ],
+ "score": 1.0,
+ "content": "computational cost.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 19.5,
+ "bbox_fs": [
+ 104,
+ 309,
+ 507,
+ 464
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 468,
+ 505,
+ 524
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 481
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 481
+ ],
+ "score": 1.0,
+ "content": "Like these prior approaches, we too address the uncertainty problem in video prediction. We propose",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 479,
+ 506,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 479,
+ 506,
+ 492
+ ],
+ "score": 1.0,
+ "content": "a general time-agnostic prediction (TAP) framework for prediction tasks. While all prior work",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 490,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 490,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "predicts at fixed time intervals, we aim to identify inherently low-uncertainty bottleneck frames with",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 502,
+ 505,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 502,
+ 505,
+ 514
+ ],
+ "score": 1.0,
+ "content": "no associated timestamp. We show how TAP may be combined with conditional GANs as well as",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 512,
+ 340,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 512,
+ 340,
+ 525
+ ],
+ "score": 1.0,
+ "content": "VAEs, to handle the residual uncertainty in its predictions.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 29,
+ "bbox_fs": [
+ 105,
+ 468,
+ 506,
+ 525
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 541,
+ 505,
+ 650
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 506,
+ 553
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 506,
+ 553
+ ],
+ "score": 1.0,
+ "content": "Bottlenecks. In hierarchical reinforcement learning, bottlenecks are proposed for discovery of",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 552,
+ 506,
+ 564
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 552,
+ 506,
+ 564
+ ],
+ "score": 1.0,
+ "content": "options (Sutton et al., 1999) in low-dimensional state spaces in (McGovern & Barto, 2001; S¸ ims¸ek",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 562,
+ 506,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 562,
+ 506,
+ 576
+ ],
+ "score": 1.0,
+ "content": "& Barto, 2009; Bacon, 2013; Metzen, 2013). Most approaches (S¸ims¸ek & Barto, 2009; Bacon,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 505,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 505,
+ 586
+ ],
+ "score": 1.0,
+ "content": "2013; Metzen, 2013) construct full transition graphs and apply notions of graph centrality to locate",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 584,
+ 505,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 584,
+ 505,
+ 598
+ ],
+ "score": 1.0,
+ "content": "bottlenecks. A multi-instance learning approach is applied in (McGovern & Barto, 2001) to mine",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 595,
+ 505,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 595,
+ 505,
+ 607
+ ],
+ "score": 1.0,
+ "content": "states that occur in successful trajectories but not in others. We consider the use of our bottleneck",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 104,
+ 606,
+ 506,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 606,
+ 506,
+ 619
+ ],
+ "score": 1.0,
+ "content": "predictions as subgoals for a hierarchical planner, which is loosely related to options in that both aim",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 630
+ ],
+ "score": 1.0,
+ "content": "to break down temporally extended trajectories into more manageable chunks. Unlike these prior",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 641
+ ],
+ "score": 1.0,
+ "content": "works, we use predictability to identify bottlenecks, and apply this to unlabeled high-dimensional",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 640,
+ 202,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 640,
+ 202,
+ 651
+ ],
+ "score": 1.0,
+ "content": "visual state trajectories.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 36.5,
+ "bbox_fs": [
+ 104,
+ 540,
+ 506,
+ 651
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 656,
+ 505,
+ 733
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 656,
+ 505,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 656,
+ 505,
+ 669
+ ],
+ "score": 1.0,
+ "content": "Concurrently with us, Neitz et al. (2018) also propose a similar idea that allows a predictor to select",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 667,
+ 505,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 667,
+ 505,
+ 680
+ ],
+ "score": 1.0,
+ "content": "when to predict, and their experiments demonstrate its advantages in specially constructed tasks with",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 678,
+ 506,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 678,
+ 506,
+ 691
+ ],
+ "score": 1.0,
+ "content": "clear bottlenecks. In comparison, we propose not just the basic time-agnostic loss (Sec 3.1), but also",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 689,
+ 505,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 689,
+ 505,
+ 702
+ ],
+ "score": 1.0,
+ "content": "improvements in Sec 3.2 through 3.5 that allow time-agnostic prediction to work in more general",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "tasks such as synthetic and real videos of robotic object manipulation. Our experiments also test the",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 711,
+ 506,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 711,
+ 506,
+ 724
+ ],
+ "score": 1.0,
+ "content": "quality of discovered bottlenecks in these scenarios and their usefulness as subgoals for hierarchical",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 104,
+ 721,
+ 147,
+ 737
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 721,
+ 147,
+ 737
+ ],
+ "score": 1.0,
+ "content": "planning.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 45,
+ "bbox_fs": [
+ 104,
+ 656,
+ 506,
+ 737
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 81,
+ 415,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 79,
+ 416,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 79,
+ 416,
+ 96
+ ],
+ "score": 1.0,
+ "content": "3 TIME-AGNOSTIC PREDICTION OF BOTTLENECK FRAMES",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 102,
+ 506,
+ 159
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 103,
+ 505,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 103,
+ 505,
+ 115
+ ],
+ "score": 1.0,
+ "content": "In visual prediction, the goal is to predict a set of unobserved target video frames given some observed",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 115,
+ 507,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 115,
+ 507,
+ 127
+ ],
+ "score": 1.0,
+ "content": "context frames. In forward prediction, the context is the first frame, and the target is all future frames.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 126,
+ 505,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 126,
+ 505,
+ 137
+ ],
+ "score": 1.0,
+ "content": "In the bidirectionally conditioned prediction case, the context is the first and the last frame, and the",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 135,
+ 506,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 135,
+ 506,
+ 150
+ ],
+ "score": 1.0,
+ "content": "frames in between are the target. In Fig 1, we may wish to predict future images of the tilting bottle,",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 147,
+ 395,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 147,
+ 395,
+ 160
+ ],
+ "score": 1.0,
+ "content": "or intermediate images of an agent who traverses the maze successfully.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 171,
+ 255,
+ 183
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 171,
+ 256,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 171,
+ 256,
+ 184
+ ],
+ "score": 1.0,
+ "content": "3.1 MINIMUM-OVER-TIME LOSS",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 190,
+ 298,
+ 289
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 190,
+ 298,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 190,
+ 298,
+ 202
+ ],
+ "score": 1.0,
+ "content": "In standard fixed-time video prediction mod-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 201,
+ 298,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 201,
+ 298,
+ 213
+ ],
+ "score": 1.0,
+ "content": "els (Ranzato et al., 2014; Oh et al., 2015; Math-",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 212,
+ 298,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 212,
+ 298,
+ 223
+ ],
+ "score": 1.0,
+ "content": "ieu et al., 2015; Vondrick et al., 2016a; Walker",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 223,
+ 298,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 223,
+ 298,
+ 235
+ ],
+ "score": 1.0,
+ "content": "et al., 2016; Finn & Levine, 2017; Xue et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 234,
+ 297,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 234,
+ 297,
+ 245
+ ],
+ "score": 1.0,
+ "content": "2016; Oh et al., 2015; Ebert et al., 2017; Finn",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 245,
+ 299,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 245,
+ 299,
+ 257
+ ],
+ "score": 1.0,
+ "content": "et al., 2016; Lee et al., 2018; Denton & Fergus,",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 255,
+ 297,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 255,
+ 171,
+ 268
+ ],
+ "score": 1.0,
+ "content": "2018), a frame",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 172,
+ 258,
+ 184,
+ 268
+ ],
+ "score": 0.87,
+ "content": "x _ { \\tau }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 184,
+ 255,
+ 275,
+ 268
+ ],
+ "score": 1.0,
+ "content": "(video frame at time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 275,
+ 258,
+ 282,
+ 267
+ ],
+ "score": 0.63,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 283,
+ 255,
+ 297,
+ 268
+ ],
+ "score": 1.0,
+ "content": ") is",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 267,
+ 297,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 267,
+ 297,
+ 279
+ ],
+ "score": 1.0,
+ "content": "selected in advance to be the training target for",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 279,
+ 298,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 279,
+ 210,
+ 290
+ ],
+ "score": 1.0,
+ "content": "some given input frames",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 211,
+ 281,
+ 217,
+ 288
+ ],
+ "score": 0.59,
+ "content": "c",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 217,
+ 279,
+ 298,
+ 290
+ ],
+ "score": 1.0,
+ "content": ". For instance, in a",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 304,
+ 203,
+ 506,
+ 265
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 304,
+ 203,
+ 506,
+ 265
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 203,
+ 506,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 203,
+ 506,
+ 265
+ ],
+ "score": 0.963,
+ "type": "image",
+ "image_path": "553b16ffacea6fe1931de7a05f2cd5a1dc88cc89a9ebb774a469ea6e63d25d18.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 17.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 304,
+ 203,
+ 506,
+ 218.5
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 304,
+ 218.5,
+ 506,
+ 234.0
+ ],
+ "spans": [],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 304,
+ 234.0,
+ 506,
+ 249.5
+ ],
+ "spans": [],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 304,
+ 249.5,
+ 506,
+ 265.0
+ ],
+ "spans": [],
+ "index": 19
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 324,
+ 272,
+ 485,
+ 284
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 322,
+ 272,
+ 486,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 322,
+ 272,
+ 486,
+ 284
+ ],
+ "score": 1.0,
+ "content": "Figure 2: The TAP minimum-over time loss.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 18.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 290,
+ 503,
+ 312
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 288,
+ 503,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 288,
+ 302,
+ 304
+ ],
+ "score": 1.0,
+ "content": "typical forward prediction setup, the input may be",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 302,
+ 291,
+ 331,
+ 300
+ ],
+ "score": 0.88,
+ "content": "c = x _ { 0 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 331,
+ 288,
+ 467,
+ 304
+ ],
+ "score": 1.0,
+ "content": ", and the target frame may be set to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 468,
+ 292,
+ 503,
+ 300
+ ],
+ "score": 0.89,
+ "content": "x _ { \\tau } = x _ { 1 }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 299,
+ 492,
+ 314
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 299,
+ 154,
+ 314
+ ],
+ "score": 1.0,
+ "content": "A predictor",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 155,
+ 301,
+ 164,
+ 310
+ ],
+ "score": 0.83,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 164,
+ 299,
+ 249,
+ 314
+ ],
+ "score": 1.0,
+ "content": "takes context frames",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 249,
+ 303,
+ 255,
+ 310
+ ],
+ "score": 0.69,
+ "content": "c",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 255,
+ 299,
+ 403,
+ 314
+ ],
+ "score": 1.0,
+ "content": "as input and produces a single frame",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 404,
+ 300,
+ 424,
+ 312
+ ],
+ "score": 0.87,
+ "content": "G ( c )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 425,
+ 299,
+ 429,
+ 314
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 429,
+ 301,
+ 438,
+ 310
+ ],
+ "score": 0.72,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 438,
+ 299,
+ 492,
+ 314
+ ],
+ "score": 1.0,
+ "content": "is trained as:",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 211,
+ 313,
+ 398,
+ 333
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 211,
+ 313,
+ 398,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 211,
+ 313,
+ 398,
+ 333
+ ],
+ "score": 0.92,
+ "content": "G ^ { * } = \\underset { G } { \\arg \\operatorname* { m i n } } \\mathcal { L } _ { 0 } ( G ) = \\underset { G } { \\arg \\operatorname* { m i n } } \\mathcal { E } ( G ( c ) , x _ { \\tau } ) ,",
+ "type": "interline_equation",
+ "image_path": "7abefb258e6867e561fc04bab08fbf27c1ebcdb5a6e54c199e2e4aa67b5643ae.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 23,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 211,
+ 313,
+ 398,
+ 333
+ ],
+ "spans": [],
+ "index": 23
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 334,
+ 505,
+ 357
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 505,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 133,
+ 347
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 335,
+ 141,
+ 344
+ ],
+ "score": 0.8,
+ "content": "\\mathcal { E }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 142,
+ 332,
+ 311,
+ 347
+ ],
+ "score": 1.0,
+ "content": "is a measure of prediction error, such as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 311,
+ 334,
+ 369,
+ 345
+ ],
+ "score": 0.87,
+ "content": "\\| G ( c ) - x _ { \\tau } \\| _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 369,
+ 332,
+ 505,
+ 347
+ ],
+ "score": 1.0,
+ "content": ".1 This predictor may be applied",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 345,
+ 466,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 345,
+ 326,
+ 358
+ ],
+ "score": 1.0,
+ "content": "recursively at test time to generate more predictions as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 326,
+ 345,
+ 418,
+ 357
+ ],
+ "score": 0.81,
+ "content": "{ \\dot { G } } ( G ( c ) ) , G ( G ( G ( c ) ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 345,
+ 466,
+ 358
+ ],
+ "score": 1.0,
+ "content": ", and so on.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 361,
+ 505,
+ 406
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 362,
+ 506,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 362,
+ 506,
+ 375
+ ],
+ "score": 1.0,
+ "content": "We propose to depart from this fixed-time paradigm by decoupling prediction from a rigid notion of",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 372,
+ 505,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 372,
+ 357,
+ 385
+ ],
+ "score": 1.0,
+ "content": "time. Instead of predicting the video frame at a specified time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 357,
+ 375,
+ 364,
+ 383
+ ],
+ "score": 0.76,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 365,
+ 372,
+ 505,
+ 385
+ ],
+ "score": 1.0,
+ "content": ", we propose to predict predictable",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 384,
+ 505,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 505,
+ 397
+ ],
+ "score": 1.0,
+ "content": "bottleneck video frames through a time-agnostic predictor (TAP), as motivated in Sec 1. To train this",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 395,
+ 370,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 395,
+ 370,
+ 407
+ ],
+ "score": 1.0,
+ "content": "predictor, we minimize the following “minimum-over-time” loss:",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 27.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 205,
+ 408,
+ 405,
+ 428
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 205,
+ 408,
+ 405,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 205,
+ 408,
+ 405,
+ 428
+ ],
+ "score": 0.91,
+ "content": "G ^ { * } = \\underset { G } { \\arg \\operatorname* { m i n } } \\ : \\mathcal { L } ( G ) = \\underset { G } { \\arg \\operatorname* { m i n } } \\ : \\underset { t \\in \\mathrm { T } } { \\operatorname* { m i n } } \\ : \\mathcal { E } ( G ( c ) , x _ { t } ) ,",
+ "type": "interline_equation",
+ "image_path": "d13da1952f7635b0f6b8307cee21cbf2e9fcb49b6c92bea2025ccd8338a670b8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 30,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 205,
+ 408,
+ 405,
+ 428
+ ],
+ "spans": [],
+ "index": 30
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 428,
+ 505,
+ 505
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 427,
+ 506,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 427,
+ 231,
+ 439
+ ],
+ "score": 1.0,
+ "content": "where the key difference from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 232,
+ 428,
+ 252,
+ 439
+ ],
+ "score": 0.26,
+ "content": "\\mathrm { E q ~ } 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 252,
+ 427,
+ 506,
+ 439
+ ],
+ "score": 1.0,
+ "content": "is that the loss is now a minimum over target times of a time-",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 438,
+ 505,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 438,
+ 164,
+ 451
+ ],
+ "score": 1.0,
+ "content": "indexed error",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 164,
+ 438,
+ 218,
+ 450
+ ],
+ "score": 0.93,
+ "content": "\\mathcal { E } _ { t } = \\mathcal { E } ( \\cdot , x _ { t } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 219,
+ 438,
+ 505,
+ 451
+ ],
+ "score": 1.0,
+ "content": ". The target times are defined by a set of time indices T. For forward",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 449,
+ 505,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 449,
+ 228,
+ 462
+ ],
+ "score": 1.0,
+ "content": "prediction starting from input",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 229,
+ 451,
+ 258,
+ 461
+ ],
+ "score": 0.9,
+ "content": "c = x _ { 0 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 258,
+ 449,
+ 350,
+ 462
+ ],
+ "score": 1.0,
+ "content": ", we may set targets to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 350,
+ 449,
+ 409,
+ 461
+ ],
+ "score": 0.93,
+ "content": "\\mathrm { T } = \\{ 1 , 2 , \\ldots \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 409,
+ 449,
+ 505,
+ 462
+ ],
+ "score": 1.0,
+ "content": ". Fig 2 depicts this idea",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 474
+ ],
+ "score": 1.0,
+ "content": "schematically. Intuitively, the penalty for a prediction is determined based on the closest matching",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 484
+ ],
+ "score": 1.0,
+ "content": "ground truth target frame. This loss incentivizes the model to latch on to “bottleneck” states in the",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 495
+ ],
+ "score": 1.0,
+ "content": "video, i.e., those with low uncertainty. In the bottle-tilting example, this would mean producing an",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 493,
+ 313,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 493,
+ 313,
+ 506
+ ],
+ "score": 1.0,
+ "content": "image of the bottle after the water has come to rest.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 510,
+ 505,
+ 599
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 523
+ ],
+ "score": 1.0,
+ "content": "One immediate concern with this minimum-over-time TAP loss might be that it could produce",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 521,
+ 505,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 521,
+ 369,
+ 533
+ ],
+ "score": 1.0,
+ "content": "degenerate predictions very close to the input conditioning frames",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 369,
+ 523,
+ 375,
+ 531
+ ],
+ "score": 0.53,
+ "content": "c",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 375,
+ 521,
+ 505,
+ 533
+ ],
+ "score": 1.0,
+ "content": ". However, as in the tilting bottle",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 532,
+ 506,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 532,
+ 506,
+ 545
+ ],
+ "score": 1.0,
+ "content": "and other cases in Fig 1, uncertainty is not always lowest closest to the observed frames. Moreover,",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 543,
+ 505,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 543,
+ 188,
+ 555
+ ],
+ "score": 1.0,
+ "content": "target frame indices",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 189,
+ 543,
+ 198,
+ 553
+ ],
+ "score": 0.47,
+ "content": "\\mathrm { T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 198,
+ 543,
+ 505,
+ 555
+ ],
+ "score": 1.0,
+ "content": "are always disjoint from the input context frames, so the model’s prediction",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 554,
+ 505,
+ 566
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 554,
+ 505,
+ 566
+ ],
+ "score": 1.0,
+ "content": "must be different from input frames by at least one step, which is no worse than the one-step-forward",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 577
+ ],
+ "score": 1.0,
+ "content": "prediction of Eq 1. In our experiments, we show cases where the minimum-over-time loss above",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 589
+ ],
+ "score": 1.0,
+ "content": "captures natural bottlenecks successfully. Further, Sec 3.2 shows how it is also possible to explicitly",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 587,
+ 271,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 261,
+ 600
+ ],
+ "score": 1.0,
+ "content": "penalize predictions near input frames",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 261,
+ 589,
+ 267,
+ 597
+ ],
+ "score": 0.41,
+ "content": "c",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 267,
+ 587,
+ 271,
+ 600
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 41.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 603,
+ 505,
+ 659
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 603,
+ 506,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 603,
+ 392,
+ 617
+ ],
+ "score": 1.0,
+ "content": "This minimum loss may be viewed as adaptively learning the time offset",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 392,
+ 606,
+ 398,
+ 614
+ ],
+ "score": 0.77,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 399,
+ 603,
+ 506,
+ 617
+ ],
+ "score": 1.0,
+ "content": ", but in fact, the predictor’s",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 614,
+ 506,
+ 627
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 614,
+ 506,
+ 627
+ ],
+ "score": 1.0,
+ "content": "task is even simpler since it is not required to provide a timestamp accompanying its prediction. For",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 639
+ ],
+ "score": 1.0,
+ "content": "example, in Fig 1(e), it need only specify which points in the maze the agent will go through; it need",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 636,
+ 506,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 636,
+ 506,
+ 650
+ ],
+ "score": 1.0,
+ "content": "not specify when. Lifting the requirement of a timestamped prediction relieves TAP approaches of a",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 106,
+ 648,
+ 217,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 648,
+ 217,
+ 661
+ ],
+ "score": 1.0,
+ "content": "significant implicit burden.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 48
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 667,
+ 504,
+ 690
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 681
+ ],
+ "score": 1.0,
+ "content": "Recursive TAP. TAP models may also be trained for recursive prediction, by minimizing the",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 678,
+ 168,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 678,
+ 168,
+ 693
+ ],
+ "score": 1.0,
+ "content": "following loss:",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ }
+ ],
+ "index": 51.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 182,
+ 689,
+ 428,
+ 714
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 182,
+ 689,
+ 428,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 182,
+ 689,
+ 428,
+ 714
+ ],
+ "score": 0.94,
+ "content": "G ^ { * } = \\underset { G } { \\arg \\operatorname* { m i n } } \\mathcal L _ { \\mathrm { r e c } } ( G ) = \\underset { G } { \\arg \\operatorname* { m i n } } \\sum _ { r } \\underset { t \\in \\mathrm { T } ( r ) } { \\min } \\mathcal { E } ( G ( c ( r ) ) , x _ { t } ) ,",
+ "type": "interline_equation",
+ "image_path": "d6c69cf1283c4a972c1f40ce26978427aaa9bcc2cf55c11ae034d55889795930.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 53,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 182,
+ 689,
+ 428,
+ 714
+ ],
+ "spans": [],
+ "index": 53
+ }
+ ]
+ }
+ ],
+ "page_idx": 2,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 118,
+ 721,
+ 455,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 719,
+ 457,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 719,
+ 457,
+ 734
+ ],
+ "score": 1.0,
+ "content": "1We omit expectations over training samples throughout this paper to avoid notational clutter.",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "score": 1.0,
+ "content": "3",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 81,
+ 415,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 79,
+ 416,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 79,
+ 416,
+ 96
+ ],
+ "score": 1.0,
+ "content": "3 TIME-AGNOSTIC PREDICTION OF BOTTLENECK FRAMES",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 102,
+ 506,
+ 159
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 103,
+ 505,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 103,
+ 505,
+ 115
+ ],
+ "score": 1.0,
+ "content": "In visual prediction, the goal is to predict a set of unobserved target video frames given some observed",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 115,
+ 507,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 115,
+ 507,
+ 127
+ ],
+ "score": 1.0,
+ "content": "context frames. In forward prediction, the context is the first frame, and the target is all future frames.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 126,
+ 505,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 126,
+ 505,
+ 137
+ ],
+ "score": 1.0,
+ "content": "In the bidirectionally conditioned prediction case, the context is the first and the last frame, and the",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 135,
+ 506,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 135,
+ 506,
+ 150
+ ],
+ "score": 1.0,
+ "content": "frames in between are the target. In Fig 1, we may wish to predict future images of the tilting bottle,",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 147,
+ 395,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 147,
+ 395,
+ 160
+ ],
+ "score": 1.0,
+ "content": "or intermediate images of an agent who traverses the maze successfully.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3,
+ "bbox_fs": [
+ 105,
+ 103,
+ 507,
+ 160
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 171,
+ 255,
+ 183
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 171,
+ 256,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 171,
+ 256,
+ 184
+ ],
+ "score": 1.0,
+ "content": "3.1 MINIMUM-OVER-TIME LOSS",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 190,
+ 298,
+ 289
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 190,
+ 298,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 190,
+ 298,
+ 202
+ ],
+ "score": 1.0,
+ "content": "In standard fixed-time video prediction mod-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 201,
+ 298,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 201,
+ 298,
+ 213
+ ],
+ "score": 1.0,
+ "content": "els (Ranzato et al., 2014; Oh et al., 2015; Math-",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 212,
+ 298,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 212,
+ 298,
+ 223
+ ],
+ "score": 1.0,
+ "content": "ieu et al., 2015; Vondrick et al., 2016a; Walker",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 223,
+ 298,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 223,
+ 298,
+ 235
+ ],
+ "score": 1.0,
+ "content": "et al., 2016; Finn & Levine, 2017; Xue et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 234,
+ 297,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 234,
+ 297,
+ 245
+ ],
+ "score": 1.0,
+ "content": "2016; Oh et al., 2015; Ebert et al., 2017; Finn",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 245,
+ 299,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 245,
+ 299,
+ 257
+ ],
+ "score": 1.0,
+ "content": "et al., 2016; Lee et al., 2018; Denton & Fergus,",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 255,
+ 297,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 255,
+ 171,
+ 268
+ ],
+ "score": 1.0,
+ "content": "2018), a frame",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 172,
+ 258,
+ 184,
+ 268
+ ],
+ "score": 0.87,
+ "content": "x _ { \\tau }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 184,
+ 255,
+ 275,
+ 268
+ ],
+ "score": 1.0,
+ "content": "(video frame at time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 275,
+ 258,
+ 282,
+ 267
+ ],
+ "score": 0.63,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 283,
+ 255,
+ 297,
+ 268
+ ],
+ "score": 1.0,
+ "content": ") is",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 267,
+ 297,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 267,
+ 297,
+ 279
+ ],
+ "score": 1.0,
+ "content": "selected in advance to be the training target for",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 279,
+ 298,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 279,
+ 210,
+ 290
+ ],
+ "score": 1.0,
+ "content": "some given input frames",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 211,
+ 281,
+ 217,
+ 288
+ ],
+ "score": 0.59,
+ "content": "c",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 217,
+ 279,
+ 298,
+ 290
+ ],
+ "score": 1.0,
+ "content": ". For instance, in a",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 11,
+ "bbox_fs": [
+ 105,
+ 190,
+ 299,
+ 290
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 304,
+ 203,
+ 506,
+ 265
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 304,
+ 203,
+ 506,
+ 265
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 203,
+ 506,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 203,
+ 506,
+ 265
+ ],
+ "score": 0.963,
+ "type": "image",
+ "image_path": "553b16ffacea6fe1931de7a05f2cd5a1dc88cc89a9ebb774a469ea6e63d25d18.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 17.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 304,
+ 203,
+ 506,
+ 218.5
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 304,
+ 218.5,
+ 506,
+ 234.0
+ ],
+ "spans": [],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 304,
+ 234.0,
+ 506,
+ 249.5
+ ],
+ "spans": [],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 304,
+ 249.5,
+ 506,
+ 265.0
+ ],
+ "spans": [],
+ "index": 19
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 324,
+ 272,
+ 485,
+ 284
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 322,
+ 272,
+ 486,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 322,
+ 272,
+ 486,
+ 284
+ ],
+ "score": 1.0,
+ "content": "Figure 2: The TAP minimum-over time loss.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 18.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 290,
+ 503,
+ 312
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 288,
+ 503,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 288,
+ 302,
+ 304
+ ],
+ "score": 1.0,
+ "content": "typical forward prediction setup, the input may be",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 302,
+ 291,
+ 331,
+ 300
+ ],
+ "score": 0.88,
+ "content": "c = x _ { 0 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 331,
+ 288,
+ 467,
+ 304
+ ],
+ "score": 1.0,
+ "content": ", and the target frame may be set to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 468,
+ 292,
+ 503,
+ 300
+ ],
+ "score": 0.89,
+ "content": "x _ { \\tau } = x _ { 1 }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 299,
+ 492,
+ 314
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 299,
+ 154,
+ 314
+ ],
+ "score": 1.0,
+ "content": "A predictor",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 155,
+ 301,
+ 164,
+ 310
+ ],
+ "score": 0.83,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 164,
+ 299,
+ 249,
+ 314
+ ],
+ "score": 1.0,
+ "content": "takes context frames",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 249,
+ 303,
+ 255,
+ 310
+ ],
+ "score": 0.69,
+ "content": "c",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 255,
+ 299,
+ 403,
+ 314
+ ],
+ "score": 1.0,
+ "content": "as input and produces a single frame",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 404,
+ 300,
+ 424,
+ 312
+ ],
+ "score": 0.87,
+ "content": "G ( c )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 425,
+ 299,
+ 429,
+ 314
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 429,
+ 301,
+ 438,
+ 310
+ ],
+ "score": 0.72,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 438,
+ 299,
+ 492,
+ 314
+ ],
+ "score": 1.0,
+ "content": "is trained as:",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21.5,
+ "bbox_fs": [
+ 105,
+ 288,
+ 503,
+ 314
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 211,
+ 313,
+ 398,
+ 333
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 211,
+ 313,
+ 398,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 211,
+ 313,
+ 398,
+ 333
+ ],
+ "score": 0.92,
+ "content": "G ^ { * } = \\underset { G } { \\arg \\operatorname* { m i n } } \\mathcal { L } _ { 0 } ( G ) = \\underset { G } { \\arg \\operatorname* { m i n } } \\mathcal { E } ( G ( c ) , x _ { \\tau } ) ,",
+ "type": "interline_equation",
+ "image_path": "7abefb258e6867e561fc04bab08fbf27c1ebcdb5a6e54c199e2e4aa67b5643ae.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 23,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 211,
+ 313,
+ 398,
+ 333
+ ],
+ "spans": [],
+ "index": 23
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 334,
+ 505,
+ 357
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 505,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 133,
+ 347
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 134,
+ 335,
+ 141,
+ 344
+ ],
+ "score": 0.8,
+ "content": "\\mathcal { E }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 142,
+ 332,
+ 311,
+ 347
+ ],
+ "score": 1.0,
+ "content": "is a measure of prediction error, such as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 311,
+ 334,
+ 369,
+ 345
+ ],
+ "score": 0.87,
+ "content": "\\| G ( c ) - x _ { \\tau } \\| _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 369,
+ 332,
+ 505,
+ 347
+ ],
+ "score": 1.0,
+ "content": ".1 This predictor may be applied",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 345,
+ 466,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 345,
+ 326,
+ 358
+ ],
+ "score": 1.0,
+ "content": "recursively at test time to generate more predictions as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 326,
+ 345,
+ 418,
+ 357
+ ],
+ "score": 0.81,
+ "content": "{ \\dot { G } } ( G ( c ) ) , G ( G ( G ( c ) ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 345,
+ 466,
+ 358
+ ],
+ "score": 1.0,
+ "content": ", and so on.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5,
+ "bbox_fs": [
+ 105,
+ 332,
+ 505,
+ 358
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 361,
+ 505,
+ 406
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 362,
+ 506,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 362,
+ 506,
+ 375
+ ],
+ "score": 1.0,
+ "content": "We propose to depart from this fixed-time paradigm by decoupling prediction from a rigid notion of",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 372,
+ 505,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 372,
+ 357,
+ 385
+ ],
+ "score": 1.0,
+ "content": "time. Instead of predicting the video frame at a specified time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 357,
+ 375,
+ 364,
+ 383
+ ],
+ "score": 0.76,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 365,
+ 372,
+ 505,
+ 385
+ ],
+ "score": 1.0,
+ "content": ", we propose to predict predictable",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 384,
+ 505,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 505,
+ 397
+ ],
+ "score": 1.0,
+ "content": "bottleneck video frames through a time-agnostic predictor (TAP), as motivated in Sec 1. To train this",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 395,
+ 370,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 395,
+ 370,
+ 407
+ ],
+ "score": 1.0,
+ "content": "predictor, we minimize the following “minimum-over-time” loss:",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 27.5,
+ "bbox_fs": [
+ 105,
+ 362,
+ 506,
+ 407
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 205,
+ 408,
+ 405,
+ 428
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 205,
+ 408,
+ 405,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 205,
+ 408,
+ 405,
+ 428
+ ],
+ "score": 0.91,
+ "content": "G ^ { * } = \\underset { G } { \\arg \\operatorname* { m i n } } \\ : \\mathcal { L } ( G ) = \\underset { G } { \\arg \\operatorname* { m i n } } \\ : \\underset { t \\in \\mathrm { T } } { \\operatorname* { m i n } } \\ : \\mathcal { E } ( G ( c ) , x _ { t } ) ,",
+ "type": "interline_equation",
+ "image_path": "d13da1952f7635b0f6b8307cee21cbf2e9fcb49b6c92bea2025ccd8338a670b8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 30,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 205,
+ 408,
+ 405,
+ 428
+ ],
+ "spans": [],
+ "index": 30
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 428,
+ 505,
+ 505
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 427,
+ 506,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 427,
+ 231,
+ 439
+ ],
+ "score": 1.0,
+ "content": "where the key difference from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 232,
+ 428,
+ 252,
+ 439
+ ],
+ "score": 0.26,
+ "content": "\\mathrm { E q ~ } 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 252,
+ 427,
+ 506,
+ 439
+ ],
+ "score": 1.0,
+ "content": "is that the loss is now a minimum over target times of a time-",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 438,
+ 505,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 438,
+ 164,
+ 451
+ ],
+ "score": 1.0,
+ "content": "indexed error",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 164,
+ 438,
+ 218,
+ 450
+ ],
+ "score": 0.93,
+ "content": "\\mathcal { E } _ { t } = \\mathcal { E } ( \\cdot , x _ { t } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 219,
+ 438,
+ 505,
+ 451
+ ],
+ "score": 1.0,
+ "content": ". The target times are defined by a set of time indices T. For forward",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 449,
+ 505,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 449,
+ 228,
+ 462
+ ],
+ "score": 1.0,
+ "content": "prediction starting from input",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 229,
+ 451,
+ 258,
+ 461
+ ],
+ "score": 0.9,
+ "content": "c = x _ { 0 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 258,
+ 449,
+ 350,
+ 462
+ ],
+ "score": 1.0,
+ "content": ", we may set targets to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 350,
+ 449,
+ 409,
+ 461
+ ],
+ "score": 0.93,
+ "content": "\\mathrm { T } = \\{ 1 , 2 , \\ldots \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 409,
+ 449,
+ 505,
+ 462
+ ],
+ "score": 1.0,
+ "content": ". Fig 2 depicts this idea",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 474
+ ],
+ "score": 1.0,
+ "content": "schematically. Intuitively, the penalty for a prediction is determined based on the closest matching",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 484
+ ],
+ "score": 1.0,
+ "content": "ground truth target frame. This loss incentivizes the model to latch on to “bottleneck” states in the",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 495
+ ],
+ "score": 1.0,
+ "content": "video, i.e., those with low uncertainty. In the bottle-tilting example, this would mean producing an",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 493,
+ 313,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 493,
+ 313,
+ 506
+ ],
+ "score": 1.0,
+ "content": "image of the bottle after the water has come to rest.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 34,
+ "bbox_fs": [
+ 105,
+ 427,
+ 506,
+ 506
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 510,
+ 505,
+ 599
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 523
+ ],
+ "score": 1.0,
+ "content": "One immediate concern with this minimum-over-time TAP loss might be that it could produce",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 521,
+ 505,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 521,
+ 369,
+ 533
+ ],
+ "score": 1.0,
+ "content": "degenerate predictions very close to the input conditioning frames",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 369,
+ 523,
+ 375,
+ 531
+ ],
+ "score": 0.53,
+ "content": "c",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 375,
+ 521,
+ 505,
+ 533
+ ],
+ "score": 1.0,
+ "content": ". However, as in the tilting bottle",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 532,
+ 506,
+ 545
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 532,
+ 506,
+ 545
+ ],
+ "score": 1.0,
+ "content": "and other cases in Fig 1, uncertainty is not always lowest closest to the observed frames. Moreover,",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 543,
+ 505,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 543,
+ 188,
+ 555
+ ],
+ "score": 1.0,
+ "content": "target frame indices",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 189,
+ 543,
+ 198,
+ 553
+ ],
+ "score": 0.47,
+ "content": "\\mathrm { T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 198,
+ 543,
+ 505,
+ 555
+ ],
+ "score": 1.0,
+ "content": "are always disjoint from the input context frames, so the model’s prediction",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 554,
+ 505,
+ 566
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 554,
+ 505,
+ 566
+ ],
+ "score": 1.0,
+ "content": "must be different from input frames by at least one step, which is no worse than the one-step-forward",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 577
+ ],
+ "score": 1.0,
+ "content": "prediction of Eq 1. In our experiments, we show cases where the minimum-over-time loss above",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 589
+ ],
+ "score": 1.0,
+ "content": "captures natural bottlenecks successfully. Further, Sec 3.2 shows how it is also possible to explicitly",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 587,
+ 271,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 261,
+ 600
+ ],
+ "score": 1.0,
+ "content": "penalize predictions near input frames",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 261,
+ 589,
+ 267,
+ 597
+ ],
+ "score": 0.41,
+ "content": "c",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 267,
+ 587,
+ 271,
+ 600
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 41.5,
+ "bbox_fs": [
+ 105,
+ 509,
+ 506,
+ 600
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 603,
+ 505,
+ 659
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 603,
+ 506,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 603,
+ 392,
+ 617
+ ],
+ "score": 1.0,
+ "content": "This minimum loss may be viewed as adaptively learning the time offset",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 392,
+ 606,
+ 398,
+ 614
+ ],
+ "score": 0.77,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 399,
+ 603,
+ 506,
+ 617
+ ],
+ "score": 1.0,
+ "content": ", but in fact, the predictor’s",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 614,
+ 506,
+ 627
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 614,
+ 506,
+ 627
+ ],
+ "score": 1.0,
+ "content": "task is even simpler since it is not required to provide a timestamp accompanying its prediction. For",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 639
+ ],
+ "score": 1.0,
+ "content": "example, in Fig 1(e), it need only specify which points in the maze the agent will go through; it need",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 636,
+ 506,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 636,
+ 506,
+ 650
+ ],
+ "score": 1.0,
+ "content": "not specify when. Lifting the requirement of a timestamped prediction relieves TAP approaches of a",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 106,
+ 648,
+ 217,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 648,
+ 217,
+ 661
+ ],
+ "score": 1.0,
+ "content": "significant implicit burden.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 48,
+ "bbox_fs": [
+ 105,
+ 603,
+ 506,
+ 661
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 667,
+ 504,
+ 690
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 681
+ ],
+ "score": 1.0,
+ "content": "Recursive TAP. TAP models may also be trained for recursive prediction, by minimizing the",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 678,
+ 168,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 678,
+ 168,
+ 693
+ ],
+ "score": 1.0,
+ "content": "following loss:",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ }
+ ],
+ "index": 51.5,
+ "bbox_fs": [
+ 105,
+ 666,
+ 505,
+ 693
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 182,
+ 689,
+ 428,
+ 714
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 182,
+ 689,
+ 428,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 182,
+ 689,
+ 428,
+ 714
+ ],
+ "score": 0.94,
+ "content": "G ^ { * } = \\underset { G } { \\arg \\operatorname* { m i n } } \\mathcal L _ { \\mathrm { r e c } } ( G ) = \\underset { G } { \\arg \\operatorname* { m i n } } \\sum _ { r } \\underset { t \\in \\mathrm { T } ( r ) } { \\min } \\mathcal { E } ( G ( c ( r ) ) , x _ { t } ) ,",
+ "type": "interline_equation",
+ "image_path": "d6c69cf1283c4a972c1f40ce26978427aaa9bcc2cf55c11ae034d55889795930.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 53,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 182,
+ 689,
+ 428,
+ 714
+ ],
+ "spans": [],
+ "index": 53
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 149
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 132,
+ 95
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 133,
+ 82,
+ 150,
+ 95
+ ],
+ "score": 0.91,
+ "content": "c ( r )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 151,
+ 82,
+ 168,
+ 95
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 169,
+ 82,
+ 189,
+ 95
+ ],
+ "score": 0.89,
+ "content": "\\mathrm { T } ( r )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 190,
+ 82,
+ 366,
+ 95
+ ],
+ "score": 1.0,
+ "content": "are the input and target set at recursion level",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 367,
+ 85,
+ 372,
+ 93
+ ],
+ "score": 0.72,
+ "content": "r",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 373,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": ", both dynamically adapted based",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 506,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 255,
+ 106
+ ],
+ "score": 1.0,
+ "content": "on the previous prediction. The input",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 255,
+ 93,
+ 273,
+ 105
+ ],
+ "score": 0.88,
+ "content": "c ( r )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 273,
+ 93,
+ 422,
+ 106
+ ],
+ "score": 1.0,
+ "content": "may be set to the previous prediction",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 423,
+ 93,
+ 472,
+ 106
+ ],
+ "score": 0.95,
+ "content": "G ( c ( r - 1 ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 472,
+ 93,
+ 506,
+ 106
+ ],
+ "score": 1.0,
+ "content": ", so that",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 103,
+ 506,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 103,
+ 225,
+ 118
+ ],
+ "score": 1.0,
+ "content": "the sequence of predictions is",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 225,
+ 105,
+ 335,
+ 117
+ ],
+ "score": 0.87,
+ "content": "( G ( c ( 0 ) ) , G ( G ( c ( 0 ) ) ) , \\dots )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 335,
+ 103,
+ 339,
+ 118
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 339,
+ 105,
+ 360,
+ 116
+ ],
+ "score": 0.76,
+ "content": "\\mathrm { T } ( r )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 360,
+ 103,
+ 506,
+ 118
+ ],
+ "score": 1.0,
+ "content": "is set to target all times after the last",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 295,
+ 127
+ ],
+ "score": 1.0,
+ "content": "prediction. In other words, if the prediction at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 295,
+ 117,
+ 320,
+ 126
+ ],
+ "score": 0.87,
+ "content": "r = 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 320,
+ 115,
+ 406,
+ 127
+ ],
+ "score": 1.0,
+ "content": "was closest to frame",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 406,
+ 117,
+ 417,
+ 127
+ ],
+ "score": 0.86,
+ "content": "x _ { 5 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 418,
+ 115,
+ 480,
+ 127
+ ],
+ "score": 1.0,
+ "content": ", the targets for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 480,
+ 116,
+ 505,
+ 126
+ ],
+ "score": 0.89,
+ "content": "r = 1",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 144,
+ 139
+ ],
+ "score": 1.0,
+ "content": "are set to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 145,
+ 126,
+ 216,
+ 138
+ ],
+ "score": 0.93,
+ "content": "\\mathrm { T } ( 1 ) = \\{ 6 , 7 , \\ldots \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 216,
+ 126,
+ 505,
+ 139
+ ],
+ "score": 1.0,
+ "content": ". While we also test recursive TAP in Sec 4, in the rest of this section, we",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 136,
+ 391,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 136,
+ 309,
+ 150
+ ],
+ "score": 1.0,
+ "content": "discuss the non-recursive formulation, building on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 309,
+ 138,
+ 329,
+ 149
+ ],
+ "score": 0.4,
+ "content": "\\operatorname { E q } 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 329,
+ 136,
+ 391,
+ 150
+ ],
+ "score": 1.0,
+ "content": ", for simplicity.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 157,
+ 506,
+ 226
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 156,
+ 507,
+ 170
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 156,
+ 507,
+ 170
+ ],
+ "score": 1.0,
+ "content": "Bidirectional TAP. Finally, while the above description of TAP has focused on forward prediction,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 167,
+ 507,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 167,
+ 446,
+ 183
+ ],
+ "score": 1.0,
+ "content": "the TAP loss of Eq 2 easily generalizes to bidirectional prediction. Given input frames",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 446,
+ 169,
+ 502,
+ 181
+ ],
+ "score": 0.91,
+ "content": "c = ( x _ { 0 } , x _ { \\mathrm { l a s t } } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 502,
+ 167,
+ 507,
+ 183
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 178,
+ 507,
+ 193
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 178,
+ 386,
+ 193
+ ],
+ "score": 1.0,
+ "content": "fixed-time bidirectional predictors might target, say, the middle frame",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 386,
+ 182,
+ 435,
+ 192
+ ],
+ "score": 0.9,
+ "content": "x _ { \\tau } = x _ { \\mathrm { l a s t / 2 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 435,
+ 178,
+ 507,
+ 193
+ ],
+ "score": 1.0,
+ "content": ". Instead, bidirec-",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 192,
+ 506,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 192,
+ 318,
+ 205
+ ],
+ "score": 1.0,
+ "content": "tional TAP models target all intermediate frames, i.e.,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 319,
+ 192,
+ 413,
+ 204
+ ],
+ "score": 0.91,
+ "content": "\\mathrm { T } = \\{ 1 , 2 , . . . , \\mathrm { l a s t } - 1 \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 413,
+ 192,
+ 506,
+ 205
+ ],
+ "score": 1.0,
+ "content": "in Eq 2. As in forward",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 202,
+ 507,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 202,
+ 507,
+ 216
+ ],
+ "score": 1.0,
+ "content": "prediction, the model has incentive to predict predictable frames. In the maze example from Fig 1,",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 213,
+ 397,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 213,
+ 397,
+ 226
+ ],
+ "score": 1.0,
+ "content": "this would mean producing an image of the agent at one of the asterisks.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 8.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 239,
+ 375,
+ 251
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 239,
+ 376,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 239,
+ 376,
+ 252
+ ],
+ "score": 1.0,
+ "content": "3.2 FROM MINIMUM TO GENERALIZED MINIMUM TAP LOSS",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 259,
+ 504,
+ 294
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 506,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 506,
+ 273
+ ],
+ "score": 1.0,
+ "content": "Within the time-agnostic prediction paradigm, we may still want to specify preferences for some",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 271,
+ 506,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 506,
+ 284
+ ],
+ "score": 1.0,
+ "content": "times over others, or for some visual properties of the predictions. Consider the minimum-over-time",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 282,
+ 383,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 282,
+ 124,
+ 294
+ ],
+ "score": 1.0,
+ "content": "loss",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 125,
+ 282,
+ 133,
+ 292
+ ],
+ "score": 0.8,
+ "content": "\\mathcal { L }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 133,
+ 282,
+ 383,
+ 294
+ ],
+ "score": 1.0,
+ "content": "in Eq 2. Taking the minimum inside, this may be rewritten as:",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 237,
+ 302,
+ 372,
+ 321
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 237,
+ 302,
+ 372,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 237,
+ 302,
+ 372,
+ 321
+ ],
+ "score": 0.93,
+ "content": "{ \\mathcal { L } } ( G ) = \\operatorname* { m i n } _ { t \\in T } { \\mathcal { E } } _ { t } = { \\mathcal { E } } _ { \\arg \\operatorname* { m i n } _ { t \\in T } { \\mathcal { E } } _ { t } } ,",
+ "type": "interline_equation",
+ "image_path": "8e2a838c3ebaff04bc9293bfbab18097ea7233e60e1c1c740e279db5c771141d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 237,
+ 302,
+ 372,
+ 321
+ ],
+ "spans": [],
+ "index": 16
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 330,
+ 504,
+ 353
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 330,
+ 505,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 330,
+ 258,
+ 343
+ ],
+ "score": 1.0,
+ "content": "where we use the time-indexed error",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 331,
+ 268,
+ 342
+ ],
+ "score": 0.88,
+ "content": "\\mathcal { E } _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 269,
+ 330,
+ 339,
+ 343
+ ],
+ "score": 1.0,
+ "content": "as shorthand for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 339,
+ 331,
+ 370,
+ 343
+ ],
+ "score": 0.92,
+ "content": "\\mathcal { E } ( . , x _ { t } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 370,
+ 330,
+ 505,
+ 343
+ ],
+ "score": 1.0,
+ "content": ". We may now extend this to the",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 341,
+ 457,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 341,
+ 457,
+ 354
+ ],
+ "score": 1.0,
+ "content": "following “generalized minimum” loss, where the outer and inner errors are decoupled:",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 17.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 256,
+ 361,
+ 354,
+ 378
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 256,
+ 361,
+ 354,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 256,
+ 361,
+ 354,
+ 378
+ ],
+ "score": 0.92,
+ "content": "\\begin{array} { r } { \\mathcal { L } ^ { \\prime } ( G ) = \\mathcal { E } _ { \\mathrm { a r g m i n } _ { t \\in T } } \\mathcal { E } _ { t } ^ { \\prime } . } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "499b1782e84de4fe899e511693d3a4ad924364cd3e84c6f7b59c9a41ff9e5202.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 19,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 256,
+ 361,
+ 354,
+ 378
+ ],
+ "spans": [],
+ "index": 19
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 386,
+ 505,
+ 464
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 385,
+ 506,
+ 400
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 385,
+ 131,
+ 400
+ ],
+ "score": 1.0,
+ "content": "Now,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 131,
+ 387,
+ 141,
+ 398
+ ],
+ "score": 0.88,
+ "content": "\\mathcal { E } _ { t } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 141,
+ 385,
+ 506,
+ 400
+ ],
+ "score": 1.0,
+ "content": ", over which the minimum is computed, could be designed to express preferences about",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 398,
+ 505,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 398,
+ 286,
+ 410
+ ],
+ "score": 1.0,
+ "content": "which frames to predict. In the simplest case,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 286,
+ 398,
+ 319,
+ 410
+ ],
+ "score": 0.93,
+ "content": "\\mathcal { E } _ { t } ^ { \\prime } = \\mathcal { E } _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 319,
+ 398,
+ 414,
+ 410
+ ],
+ "score": 1.0,
+ "content": ", and the loss reduces to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 414,
+ 398,
+ 434,
+ 409
+ ],
+ "score": 0.42,
+ "content": "\\operatorname { E q } 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 434,
+ 398,
+ 505,
+ 410
+ ],
+ "score": 1.0,
+ "content": ". Instead, suppose",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 408,
+ 506,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 344,
+ 421
+ ],
+ "score": 1.0,
+ "content": "that predictions at some times are preferred over others. Let",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 409,
+ 364,
+ 421
+ ],
+ "score": 0.92,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 364,
+ 408,
+ 506,
+ 421
+ ],
+ "score": 1.0,
+ "content": "express the preference value for all",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 505,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 156,
+ 433
+ ],
+ "score": 1.0,
+ "content": "target times",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 156,
+ 420,
+ 180,
+ 430
+ ],
+ "score": 0.89,
+ "content": "t \\in T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 419,
+ 241,
+ 433
+ ],
+ "score": 1.0,
+ "content": ", so that higher",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 242,
+ 420,
+ 261,
+ 432
+ ],
+ "score": 0.91,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 262,
+ 419,
+ 449,
+ 433
+ ],
+ "score": 1.0,
+ "content": "indicates higher preference. Then we may set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 450,
+ 420,
+ 505,
+ 432
+ ],
+ "score": 0.91,
+ "content": "\\mathcal { E } _ { t } ^ { \\prime } = \\mathcal { E } _ { t } / w ( t )",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 430,
+ 506,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 430,
+ 161,
+ 443
+ ],
+ "score": 1.0,
+ "content": "so that times",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 162,
+ 432,
+ 167,
+ 441
+ ],
+ "score": 0.72,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 167,
+ 430,
+ 218,
+ 443
+ ],
+ "score": 1.0,
+ "content": "with higher",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 218,
+ 431,
+ 238,
+ 442
+ ],
+ "score": 0.91,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 238,
+ 430,
+ 473,
+ 443
+ ],
+ "score": 1.0,
+ "content": "are preferred in the arg min. In our experiments, we set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 473,
+ 432,
+ 493,
+ 443
+ ],
+ "score": 0.9,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 493,
+ 430,
+ 506,
+ 443
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 441,
+ 506,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 441,
+ 506,
+ 454
+ ],
+ "score": 1.0,
+ "content": "linearly increase with time during forward prediction and to a truncated discrete Gaussian centered at",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 452,
+ 268,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 452,
+ 268,
+ 465
+ ],
+ "score": 1.0,
+ "content": "the midpoint in bidirectional prediction.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 469,
+ 505,
+ 536
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 470,
+ 505,
+ 481
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 470,
+ 505,
+ 481
+ ],
+ "score": 1.0,
+ "content": "At this point, one might ask: could we not directly incorporate preferences into the outer error?",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 480,
+ 506,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 480,
+ 271,
+ 493
+ ],
+ "score": 1.0,
+ "content": "For instance, why not simply optimize",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 272,
+ 480,
+ 329,
+ 492
+ ],
+ "score": 0.92,
+ "content": "\\mathrm { m i n } _ { t } \\mathcal { E } _ { t } / w ( t ) ^ { \\star }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 329,
+ 480,
+ 506,
+ 493
+ ],
+ "score": 1.0,
+ "content": "? Unfortunately, that would have the side-",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 491,
+ 505,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 491,
+ 453,
+ 504
+ ],
+ "score": 1.0,
+ "content": "effect of downweighting the errors computed against frames with higher preferences",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 453,
+ 491,
+ 473,
+ 504
+ ],
+ "score": 0.91,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 474,
+ 491,
+ 505,
+ 504
+ ],
+ "score": 1.0,
+ "content": ", which",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 104,
+ 501,
+ 506,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 501,
+ 412,
+ 516
+ ],
+ "score": 1.0,
+ "content": "is counterproductive. Decoupling the outer and inner errors instead, as in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 412,
+ 502,
+ 433,
+ 514
+ ],
+ "score": 0.36,
+ "content": "\\operatorname { E q } 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 433,
+ 501,
+ 506,
+ 516
+ ],
+ "score": 1.0,
+ "content": ", allows applying",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 513,
+ 506,
+ 526
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 513,
+ 154,
+ 526
+ ],
+ "score": 1.0,
+ "content": "preferences",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 154,
+ 513,
+ 174,
+ 525
+ ],
+ "score": 0.91,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 174,
+ 513,
+ 506,
+ 526
+ ],
+ "score": 1.0,
+ "content": "only to select the target frame to compute the outer loss against; the outer loss itself",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 524,
+ 407,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 407,
+ 536
+ ],
+ "score": 1.0,
+ "content": "penalizes prediction errors equally regardless of which frame was selected.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 541,
+ 505,
+ 586
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 541,
+ 505,
+ 553
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 541,
+ 505,
+ 553
+ ],
+ "score": 1.0,
+ "content": "The generalized minimum formulation may be used to express other kinds of preferences too. For",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 553,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 553,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "instance, when using predictions as subgoals in a planner, perhaps some states are more expensive to",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 564,
+ 505,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 505,
+ 576
+ ],
+ "score": 1.0,
+ "content": "reach than others. We also use the generalized minimum to select frames using different criteria than",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 574,
+ 339,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 574,
+ 339,
+ 586
+ ],
+ "score": 1.0,
+ "content": "the prediction penalty itself, as we will discuss in Sec 3.5.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 34.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 599,
+ 298,
+ 611
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 598,
+ 299,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 598,
+ 299,
+ 613
+ ],
+ "score": 1.0,
+ "content": "3.3 TIME-AGNOSTIC CONDITIONAL GANS",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 619,
+ 505,
+ 686
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 619,
+ 506,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 619,
+ 281,
+ 632
+ ],
+ "score": 1.0,
+ "content": "TAP is not limited to simple losses such as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 282,
+ 620,
+ 292,
+ 631
+ ],
+ "score": 0.88,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 292,
+ 619,
+ 304,
+ 632
+ ],
+ "score": 1.0,
+ "content": "or",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 620,
+ 314,
+ 631
+ ],
+ "score": 0.88,
+ "content": "\\ell _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 314,
+ 619,
+ 506,
+ 632
+ ],
+ "score": 1.0,
+ "content": "errors; it can be extended to handle expressive",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 631,
+ 505,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 631,
+ 505,
+ 644
+ ],
+ "score": 1.0,
+ "content": "GAN losses to improve perceptual quality. A standard conditional GAN (CGAN) in fixed-time",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 641,
+ 506,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 641,
+ 236,
+ 654
+ ],
+ "score": 1.0,
+ "content": "video prediction targeting time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 644,
+ 243,
+ 651
+ ],
+ "score": 0.77,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 244,
+ 641,
+ 420,
+ 654
+ ],
+ "score": 1.0,
+ "content": "works as follows: given a “discriminator”",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 420,
+ 642,
+ 430,
+ 651
+ ],
+ "score": 0.82,
+ "content": "D",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 430,
+ 641,
+ 506,
+ 654
+ ],
+ "score": 1.0,
+ "content": "that outputs 0 for",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 653,
+ 505,
+ 665
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 653,
+ 408,
+ 665
+ ],
+ "score": 1.0,
+ "content": "input-prediction tuples and 1 for input-ground truth tuples, the generator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 408,
+ 653,
+ 417,
+ 663
+ ],
+ "score": 0.82,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 417,
+ 653,
+ 505,
+ 665
+ ],
+ "score": 1.0,
+ "content": "is trained to fool the",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 663,
+ 507,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 663,
+ 507,
+ 676
+ ],
+ "score": 1.0,
+ "content": "discriminator. The discriminator in turn is trained adversarially using a binary cross-entropy loss.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 675,
+ 247,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 675,
+ 247,
+ 687
+ ],
+ "score": 1.0,
+ "content": "The CGAN objective is written as:",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 40.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 198,
+ 694,
+ 412,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 198,
+ 694,
+ 412,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 198,
+ 694,
+ 412,
+ 732
+ ],
+ "score": 0.9,
+ "content": "\\begin{array} { r l } & { G ^ { * } = \\underset { G } { \\arg \\operatorname* { m i n } } \\ \\underset { D } { \\operatorname* { m a x } } \\mathcal { L } _ { \\mathrm { c g a n } } ( G , D ) , } \\\\ & { } \\\\ & { \\mathcal { L } _ { \\mathrm { c g a n } } ( G , D ) = \\log ( D ( c , x _ { \\tau } ) ) + \\log ( 1 - D ( c , G ( c ) ) } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "a344bd70cbb79bd7a16342e2a26baaad21507932570028252c1ff0aaf86ac95f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 44.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 198,
+ 694,
+ 412,
+ 713.0
+ ],
+ "spans": [],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 198,
+ 713.0,
+ 412,
+ 732.0
+ ],
+ "spans": [],
+ "index": 45
+ }
+ ]
+ }
+ ],
+ "page_idx": 3,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 759
+ ],
+ "lines": []
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 149
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 132,
+ 95
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 133,
+ 82,
+ 150,
+ 95
+ ],
+ "score": 0.91,
+ "content": "c ( r )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 151,
+ 82,
+ 168,
+ 95
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 169,
+ 82,
+ 189,
+ 95
+ ],
+ "score": 0.89,
+ "content": "\\mathrm { T } ( r )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 190,
+ 82,
+ 366,
+ 95
+ ],
+ "score": 1.0,
+ "content": "are the input and target set at recursion level",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 367,
+ 85,
+ 372,
+ 93
+ ],
+ "score": 0.72,
+ "content": "r",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 373,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": ", both dynamically adapted based",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 506,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 255,
+ 106
+ ],
+ "score": 1.0,
+ "content": "on the previous prediction. The input",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 255,
+ 93,
+ 273,
+ 105
+ ],
+ "score": 0.88,
+ "content": "c ( r )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 273,
+ 93,
+ 422,
+ 106
+ ],
+ "score": 1.0,
+ "content": "may be set to the previous prediction",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 423,
+ 93,
+ 472,
+ 106
+ ],
+ "score": 0.95,
+ "content": "G ( c ( r - 1 ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 472,
+ 93,
+ 506,
+ 106
+ ],
+ "score": 1.0,
+ "content": ", so that",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 103,
+ 506,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 103,
+ 225,
+ 118
+ ],
+ "score": 1.0,
+ "content": "the sequence of predictions is",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 225,
+ 105,
+ 335,
+ 117
+ ],
+ "score": 0.87,
+ "content": "( G ( c ( 0 ) ) , G ( G ( c ( 0 ) ) ) , \\dots )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 335,
+ 103,
+ 339,
+ 118
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 339,
+ 105,
+ 360,
+ 116
+ ],
+ "score": 0.76,
+ "content": "\\mathrm { T } ( r )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 360,
+ 103,
+ 506,
+ 118
+ ],
+ "score": 1.0,
+ "content": "is set to target all times after the last",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 295,
+ 127
+ ],
+ "score": 1.0,
+ "content": "prediction. In other words, if the prediction at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 295,
+ 117,
+ 320,
+ 126
+ ],
+ "score": 0.87,
+ "content": "r = 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 320,
+ 115,
+ 406,
+ 127
+ ],
+ "score": 1.0,
+ "content": "was closest to frame",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 406,
+ 117,
+ 417,
+ 127
+ ],
+ "score": 0.86,
+ "content": "x _ { 5 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 418,
+ 115,
+ 480,
+ 127
+ ],
+ "score": 1.0,
+ "content": ", the targets for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 480,
+ 116,
+ 505,
+ 126
+ ],
+ "score": 0.89,
+ "content": "r = 1",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 144,
+ 139
+ ],
+ "score": 1.0,
+ "content": "are set to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 145,
+ 126,
+ 216,
+ 138
+ ],
+ "score": 0.93,
+ "content": "\\mathrm { T } ( 1 ) = \\{ 6 , 7 , \\ldots \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 216,
+ 126,
+ 505,
+ 139
+ ],
+ "score": 1.0,
+ "content": ". While we also test recursive TAP in Sec 4, in the rest of this section, we",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 136,
+ 391,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 136,
+ 309,
+ 150
+ ],
+ "score": 1.0,
+ "content": "discuss the non-recursive formulation, building on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 309,
+ 138,
+ 329,
+ 149
+ ],
+ "score": 0.4,
+ "content": "\\operatorname { E q } 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 329,
+ 136,
+ 391,
+ 150
+ ],
+ "score": 1.0,
+ "content": ", for simplicity.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 2.5,
+ "bbox_fs": [
+ 105,
+ 82,
+ 506,
+ 150
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 157,
+ 506,
+ 226
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 156,
+ 507,
+ 170
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 156,
+ 507,
+ 170
+ ],
+ "score": 1.0,
+ "content": "Bidirectional TAP. Finally, while the above description of TAP has focused on forward prediction,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 167,
+ 507,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 167,
+ 446,
+ 183
+ ],
+ "score": 1.0,
+ "content": "the TAP loss of Eq 2 easily generalizes to bidirectional prediction. Given input frames",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 446,
+ 169,
+ 502,
+ 181
+ ],
+ "score": 0.91,
+ "content": "c = ( x _ { 0 } , x _ { \\mathrm { l a s t } } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 502,
+ 167,
+ 507,
+ 183
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 178,
+ 507,
+ 193
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 178,
+ 386,
+ 193
+ ],
+ "score": 1.0,
+ "content": "fixed-time bidirectional predictors might target, say, the middle frame",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 386,
+ 182,
+ 435,
+ 192
+ ],
+ "score": 0.9,
+ "content": "x _ { \\tau } = x _ { \\mathrm { l a s t / 2 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 435,
+ 178,
+ 507,
+ 193
+ ],
+ "score": 1.0,
+ "content": ". Instead, bidirec-",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 192,
+ 506,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 192,
+ 318,
+ 205
+ ],
+ "score": 1.0,
+ "content": "tional TAP models target all intermediate frames, i.e.,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 319,
+ 192,
+ 413,
+ 204
+ ],
+ "score": 0.91,
+ "content": "\\mathrm { T } = \\{ 1 , 2 , . . . , \\mathrm { l a s t } - 1 \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 413,
+ 192,
+ 506,
+ 205
+ ],
+ "score": 1.0,
+ "content": "in Eq 2. As in forward",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 202,
+ 507,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 202,
+ 507,
+ 216
+ ],
+ "score": 1.0,
+ "content": "prediction, the model has incentive to predict predictable frames. In the maze example from Fig 1,",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 213,
+ 397,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 213,
+ 397,
+ 226
+ ],
+ "score": 1.0,
+ "content": "this would mean producing an image of the agent at one of the asterisks.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 8.5,
+ "bbox_fs": [
+ 105,
+ 156,
+ 507,
+ 226
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 239,
+ 375,
+ 251
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 239,
+ 376,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 239,
+ 376,
+ 252
+ ],
+ "score": 1.0,
+ "content": "3.2 FROM MINIMUM TO GENERALIZED MINIMUM TAP LOSS",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 259,
+ 504,
+ 294
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 506,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 506,
+ 273
+ ],
+ "score": 1.0,
+ "content": "Within the time-agnostic prediction paradigm, we may still want to specify preferences for some",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 271,
+ 506,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 506,
+ 284
+ ],
+ "score": 1.0,
+ "content": "times over others, or for some visual properties of the predictions. Consider the minimum-over-time",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 282,
+ 383,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 282,
+ 124,
+ 294
+ ],
+ "score": 1.0,
+ "content": "loss",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 125,
+ 282,
+ 133,
+ 292
+ ],
+ "score": 0.8,
+ "content": "\\mathcal { L }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 133,
+ 282,
+ 383,
+ 294
+ ],
+ "score": 1.0,
+ "content": "in Eq 2. Taking the minimum inside, this may be rewritten as:",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 14,
+ "bbox_fs": [
+ 105,
+ 259,
+ 506,
+ 294
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 237,
+ 302,
+ 372,
+ 321
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 237,
+ 302,
+ 372,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 237,
+ 302,
+ 372,
+ 321
+ ],
+ "score": 0.93,
+ "content": "{ \\mathcal { L } } ( G ) = \\operatorname* { m i n } _ { t \\in T } { \\mathcal { E } } _ { t } = { \\mathcal { E } } _ { \\arg \\operatorname* { m i n } _ { t \\in T } { \\mathcal { E } } _ { t } } ,",
+ "type": "interline_equation",
+ "image_path": "8e2a838c3ebaff04bc9293bfbab18097ea7233e60e1c1c740e279db5c771141d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 237,
+ 302,
+ 372,
+ 321
+ ],
+ "spans": [],
+ "index": 16
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 330,
+ 504,
+ 353
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 330,
+ 505,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 330,
+ 258,
+ 343
+ ],
+ "score": 1.0,
+ "content": "where we use the time-indexed error",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 331,
+ 268,
+ 342
+ ],
+ "score": 0.88,
+ "content": "\\mathcal { E } _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 269,
+ 330,
+ 339,
+ 343
+ ],
+ "score": 1.0,
+ "content": "as shorthand for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 339,
+ 331,
+ 370,
+ 343
+ ],
+ "score": 0.92,
+ "content": "\\mathcal { E } ( . , x _ { t } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 370,
+ 330,
+ 505,
+ 343
+ ],
+ "score": 1.0,
+ "content": ". We may now extend this to the",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 341,
+ 457,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 341,
+ 457,
+ 354
+ ],
+ "score": 1.0,
+ "content": "following “generalized minimum” loss, where the outer and inner errors are decoupled:",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 17.5,
+ "bbox_fs": [
+ 106,
+ 330,
+ 505,
+ 354
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 256,
+ 361,
+ 354,
+ 378
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 256,
+ 361,
+ 354,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 256,
+ 361,
+ 354,
+ 378
+ ],
+ "score": 0.92,
+ "content": "\\begin{array} { r } { \\mathcal { L } ^ { \\prime } ( G ) = \\mathcal { E } _ { \\mathrm { a r g m i n } _ { t \\in T } } \\mathcal { E } _ { t } ^ { \\prime } . } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "499b1782e84de4fe899e511693d3a4ad924364cd3e84c6f7b59c9a41ff9e5202.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 19,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 256,
+ 361,
+ 354,
+ 378
+ ],
+ "spans": [],
+ "index": 19
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 386,
+ 505,
+ 464
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 385,
+ 506,
+ 400
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 385,
+ 131,
+ 400
+ ],
+ "score": 1.0,
+ "content": "Now,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 131,
+ 387,
+ 141,
+ 398
+ ],
+ "score": 0.88,
+ "content": "\\mathcal { E } _ { t } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 141,
+ 385,
+ 506,
+ 400
+ ],
+ "score": 1.0,
+ "content": ", over which the minimum is computed, could be designed to express preferences about",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 398,
+ 505,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 398,
+ 286,
+ 410
+ ],
+ "score": 1.0,
+ "content": "which frames to predict. In the simplest case,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 286,
+ 398,
+ 319,
+ 410
+ ],
+ "score": 0.93,
+ "content": "\\mathcal { E } _ { t } ^ { \\prime } = \\mathcal { E } _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 319,
+ 398,
+ 414,
+ 410
+ ],
+ "score": 1.0,
+ "content": ", and the loss reduces to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 414,
+ 398,
+ 434,
+ 409
+ ],
+ "score": 0.42,
+ "content": "\\operatorname { E q } 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 434,
+ 398,
+ 505,
+ 410
+ ],
+ "score": 1.0,
+ "content": ". Instead, suppose",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 408,
+ 506,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 344,
+ 421
+ ],
+ "score": 1.0,
+ "content": "that predictions at some times are preferred over others. Let",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 409,
+ 364,
+ 421
+ ],
+ "score": 0.92,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 364,
+ 408,
+ 506,
+ 421
+ ],
+ "score": 1.0,
+ "content": "express the preference value for all",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 505,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 156,
+ 433
+ ],
+ "score": 1.0,
+ "content": "target times",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 156,
+ 420,
+ 180,
+ 430
+ ],
+ "score": 0.89,
+ "content": "t \\in T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 419,
+ 241,
+ 433
+ ],
+ "score": 1.0,
+ "content": ", so that higher",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 242,
+ 420,
+ 261,
+ 432
+ ],
+ "score": 0.91,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 262,
+ 419,
+ 449,
+ 433
+ ],
+ "score": 1.0,
+ "content": "indicates higher preference. Then we may set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 450,
+ 420,
+ 505,
+ 432
+ ],
+ "score": 0.91,
+ "content": "\\mathcal { E } _ { t } ^ { \\prime } = \\mathcal { E } _ { t } / w ( t )",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 430,
+ 506,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 430,
+ 161,
+ 443
+ ],
+ "score": 1.0,
+ "content": "so that times",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 162,
+ 432,
+ 167,
+ 441
+ ],
+ "score": 0.72,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 167,
+ 430,
+ 218,
+ 443
+ ],
+ "score": 1.0,
+ "content": "with higher",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 218,
+ 431,
+ 238,
+ 442
+ ],
+ "score": 0.91,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 238,
+ 430,
+ 473,
+ 443
+ ],
+ "score": 1.0,
+ "content": "are preferred in the arg min. In our experiments, we set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 473,
+ 432,
+ 493,
+ 443
+ ],
+ "score": 0.9,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 493,
+ 430,
+ 506,
+ 443
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 441,
+ 506,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 441,
+ 506,
+ 454
+ ],
+ "score": 1.0,
+ "content": "linearly increase with time during forward prediction and to a truncated discrete Gaussian centered at",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 452,
+ 268,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 452,
+ 268,
+ 465
+ ],
+ "score": 1.0,
+ "content": "the midpoint in bidirectional prediction.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 23,
+ "bbox_fs": [
+ 105,
+ 385,
+ 506,
+ 465
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 469,
+ 505,
+ 536
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 470,
+ 505,
+ 481
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 470,
+ 505,
+ 481
+ ],
+ "score": 1.0,
+ "content": "At this point, one might ask: could we not directly incorporate preferences into the outer error?",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 480,
+ 506,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 480,
+ 271,
+ 493
+ ],
+ "score": 1.0,
+ "content": "For instance, why not simply optimize",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 272,
+ 480,
+ 329,
+ 492
+ ],
+ "score": 0.92,
+ "content": "\\mathrm { m i n } _ { t } \\mathcal { E } _ { t } / w ( t ) ^ { \\star }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 329,
+ 480,
+ 506,
+ 493
+ ],
+ "score": 1.0,
+ "content": "? Unfortunately, that would have the side-",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 491,
+ 505,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 491,
+ 453,
+ 504
+ ],
+ "score": 1.0,
+ "content": "effect of downweighting the errors computed against frames with higher preferences",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 453,
+ 491,
+ 473,
+ 504
+ ],
+ "score": 0.91,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 474,
+ 491,
+ 505,
+ 504
+ ],
+ "score": 1.0,
+ "content": ", which",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 104,
+ 501,
+ 506,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 501,
+ 412,
+ 516
+ ],
+ "score": 1.0,
+ "content": "is counterproductive. Decoupling the outer and inner errors instead, as in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 412,
+ 502,
+ 433,
+ 514
+ ],
+ "score": 0.36,
+ "content": "\\operatorname { E q } 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 433,
+ 501,
+ 506,
+ 516
+ ],
+ "score": 1.0,
+ "content": ", allows applying",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 513,
+ 506,
+ 526
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 513,
+ 154,
+ 526
+ ],
+ "score": 1.0,
+ "content": "preferences",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 154,
+ 513,
+ 174,
+ 525
+ ],
+ "score": 0.91,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 174,
+ 513,
+ 506,
+ 526
+ ],
+ "score": 1.0,
+ "content": "only to select the target frame to compute the outer loss against; the outer loss itself",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 524,
+ 407,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 407,
+ 536
+ ],
+ "score": 1.0,
+ "content": "penalizes prediction errors equally regardless of which frame was selected.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 29.5,
+ "bbox_fs": [
+ 104,
+ 470,
+ 506,
+ 536
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 541,
+ 505,
+ 586
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 541,
+ 505,
+ 553
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 541,
+ 505,
+ 553
+ ],
+ "score": 1.0,
+ "content": "The generalized minimum formulation may be used to express other kinds of preferences too. For",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 553,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 553,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "instance, when using predictions as subgoals in a planner, perhaps some states are more expensive to",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 564,
+ 505,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 564,
+ 505,
+ 576
+ ],
+ "score": 1.0,
+ "content": "reach than others. We also use the generalized minimum to select frames using different criteria than",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 574,
+ 339,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 574,
+ 339,
+ 586
+ ],
+ "score": 1.0,
+ "content": "the prediction penalty itself, as we will discuss in Sec 3.5.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 34.5,
+ "bbox_fs": [
+ 105,
+ 541,
+ 505,
+ 586
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 599,
+ 298,
+ 611
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 598,
+ 299,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 598,
+ 299,
+ 613
+ ],
+ "score": 1.0,
+ "content": "3.3 TIME-AGNOSTIC CONDITIONAL GANS",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 619,
+ 505,
+ 686
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 619,
+ 506,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 619,
+ 281,
+ 632
+ ],
+ "score": 1.0,
+ "content": "TAP is not limited to simple losses such as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 282,
+ 620,
+ 292,
+ 631
+ ],
+ "score": 0.88,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 292,
+ 619,
+ 304,
+ 632
+ ],
+ "score": 1.0,
+ "content": "or",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 620,
+ 314,
+ 631
+ ],
+ "score": 0.88,
+ "content": "\\ell _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 314,
+ 619,
+ 506,
+ 632
+ ],
+ "score": 1.0,
+ "content": "errors; it can be extended to handle expressive",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 631,
+ 505,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 631,
+ 505,
+ 644
+ ],
+ "score": 1.0,
+ "content": "GAN losses to improve perceptual quality. A standard conditional GAN (CGAN) in fixed-time",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 641,
+ 506,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 641,
+ 236,
+ 654
+ ],
+ "score": 1.0,
+ "content": "video prediction targeting time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 644,
+ 243,
+ 651
+ ],
+ "score": 0.77,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 244,
+ 641,
+ 420,
+ 654
+ ],
+ "score": 1.0,
+ "content": "works as follows: given a “discriminator”",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 420,
+ 642,
+ 430,
+ 651
+ ],
+ "score": 0.82,
+ "content": "D",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 430,
+ 641,
+ 506,
+ 654
+ ],
+ "score": 1.0,
+ "content": "that outputs 0 for",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 653,
+ 505,
+ 665
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 653,
+ 408,
+ 665
+ ],
+ "score": 1.0,
+ "content": "input-prediction tuples and 1 for input-ground truth tuples, the generator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 408,
+ 653,
+ 417,
+ 663
+ ],
+ "score": 0.82,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 417,
+ 653,
+ 505,
+ 665
+ ],
+ "score": 1.0,
+ "content": "is trained to fool the",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 663,
+ 507,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 663,
+ 507,
+ 676
+ ],
+ "score": 1.0,
+ "content": "discriminator. The discriminator in turn is trained adversarially using a binary cross-entropy loss.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 675,
+ 247,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 675,
+ 247,
+ 687
+ ],
+ "score": 1.0,
+ "content": "The CGAN objective is written as:",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 105,
+ 619,
+ 507,
+ 687
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 198,
+ 694,
+ 412,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 198,
+ 694,
+ 412,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 198,
+ 694,
+ 412,
+ 732
+ ],
+ "score": 0.9,
+ "content": "\\begin{array} { r l } & { G ^ { * } = \\underset { G } { \\arg \\operatorname* { m i n } } \\ \\underset { D } { \\operatorname* { m a x } } \\mathcal { L } _ { \\mathrm { c g a n } } ( G , D ) , } \\\\ & { } \\\\ & { \\mathcal { L } _ { \\mathrm { c g a n } } ( G , D ) = \\log ( D ( c , x _ { \\tau } ) ) + \\log ( 1 - D ( c , G ( c ) ) } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "a344bd70cbb79bd7a16342e2a26baaad21507932570028252c1ff0aaf86ac95f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 44.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 198,
+ 694,
+ 412,
+ 713.0
+ ],
+ "spans": [],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 198,
+ 713.0,
+ 412,
+ 732.0
+ ],
+ "spans": [],
+ "index": 45
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 81,
+ 505,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 506,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 295,
+ 96
+ ],
+ "score": 1.0,
+ "content": "To make this compatible with TAP, we train",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 296,
+ 83,
+ 310,
+ 95
+ ],
+ "score": 0.86,
+ "content": "| \\mathrm { T } |",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 310,
+ 81,
+ 374,
+ 96
+ ],
+ "score": 1.0,
+ "content": "discriminators",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 374,
+ 82,
+ 397,
+ 95
+ ],
+ "score": 0.92,
+ "content": "\\{ D _ { t } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 397,
+ 81,
+ 506,
+ 96
+ ],
+ "score": 1.0,
+ "content": ", one per timestep. Then,",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 359,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 159,
+ 106
+ ],
+ "score": 1.0,
+ "content": "analogous to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 159,
+ 94,
+ 180,
+ 105
+ ],
+ "score": 0.28,
+ "content": "\\operatorname { E q } 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 180,
+ 93,
+ 359,
+ 106
+ ],
+ "score": 1.0,
+ "content": ", we may define a time-agnostic CGAN loss:",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 141,
+ 111,
+ 469,
+ 163
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 111,
+ 469,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 111,
+ 469,
+ 163
+ ],
+ "score": 0.93,
+ "content": "\\begin{array} { r l } & { G ^ { * } = \\underset { G } { \\arg \\operatorname* { m i n } } \\underset { t \\in T } { \\operatorname* { m i n } } \\underset { D _ { t } } { \\operatorname* { m a x } } \\mathcal { L } _ { \\mathrm { c g a n } } ^ { t } ( G , D _ { t } ) , } \\\\ & { \\mathcal { L } _ { \\mathrm { c g a n } } ^ { t } ( G , D _ { t } ) = \\log D _ { t } ( c , x _ { t } ) + \\log ( 1 - D _ { t } ( c , G ( c ) ) ) + \\underset { t ^ { \\prime } \\neq t } { \\sum } \\log ( 1 - D _ { t } ( c , x _ { t ^ { \\prime } } ) ) , } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "642032ec3b2c5e42befe78ad2d47ede55bef3f7c187593e8ad908907a0646f0a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 141,
+ 111,
+ 469,
+ 128.33333333333334
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 141,
+ 128.33333333333334,
+ 469,
+ 145.66666666666669
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 141,
+ 145.66666666666669,
+ 469,
+ 163.00000000000003
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 169,
+ 506,
+ 236
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 169,
+ 505,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 169,
+ 127,
+ 182
+ ],
+ "score": 1.0,
+ "content": "Like",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 128,
+ 170,
+ 173,
+ 182
+ ],
+ "score": 0.42,
+ "content": "\\mathrm { E q } 6 , \\mathrm { E q } 7",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 173,
+ 169,
+ 441,
+ 182
+ ],
+ "score": 1.0,
+ "content": "defines a cross-entropy loss. The first two terms are analogous to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 441,
+ 170,
+ 462,
+ 182
+ ],
+ "score": 0.37,
+ "content": "\\operatorname { E q } 6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 463,
+ 169,
+ 505,
+ 182
+ ],
+ "score": 1.0,
+ "content": "— for the",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 107,
+ 180,
+ 505,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 182,
+ 111,
+ 191
+ ],
+ "score": 0.75,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 112,
+ 180,
+ 198,
+ 194
+ ],
+ "score": 1.0,
+ "content": "-th discriminator, the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 198,
+ 182,
+ 203,
+ 191
+ ],
+ "score": 0.76,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 180,
+ 505,
+ 194
+ ],
+ "score": 1.0,
+ "content": "-th frame provides a positive, and the generated frame provides a negative",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 191,
+ 507,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 191,
+ 430,
+ 205
+ ],
+ "score": 1.0,
+ "content": "instance. The third term treats ground truth video frames occurring at other times",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 431,
+ 193,
+ 453,
+ 203
+ ],
+ "score": 0.85,
+ "content": "x _ { t ^ { \\prime } \\neq t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 453,
+ 191,
+ 507,
+ 205
+ ],
+ "score": 1.0,
+ "content": "as negatives.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 202,
+ 506,
+ 215
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 202,
+ 332,
+ 215
+ ],
+ "score": 1.0,
+ "content": "In practice, we train a single discriminator network with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 332,
+ 203,
+ 346,
+ 215
+ ],
+ "score": 0.9,
+ "content": "| T |",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 346,
+ 202,
+ 430,
+ 215
+ ],
+ "score": 1.0,
+ "content": "outputs that serve as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 430,
+ 204,
+ 453,
+ 215
+ ],
+ "score": 0.9,
+ "content": "\\{ D _ { t } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 453,
+ 202,
+ 506,
+ 215
+ ],
+ "score": 1.0,
+ "content": ". Further, for",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 214,
+ 505,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 214,
+ 358,
+ 226
+ ],
+ "score": 1.0,
+ "content": "computational efficiency, we approximate the summation over",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 359,
+ 214,
+ 383,
+ 225
+ ],
+ "score": 0.91,
+ "content": "t ^ { \\prime } \\neq t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 383,
+ 214,
+ 505,
+ 226
+ ],
+ "score": 1.0,
+ "content": "by sampling a single frame at",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 225,
+ 463,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 225,
+ 463,
+ 237
+ ],
+ "score": 1.0,
+ "content": "random for each training video at each iteration. Appendix A provides additional details.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 249,
+ 296,
+ 261
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 249,
+ 297,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 249,
+ 297,
+ 261
+ ],
+ "score": 1.0,
+ "content": "3.4 TIME-AGNOSTIC CONDITIONAL VAES",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 269,
+ 505,
+ 336
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 270,
+ 505,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 270,
+ 505,
+ 281
+ ],
+ "score": 1.0,
+ "content": "While TAP targets low-uncertainty bottleneck states, it may be integrated with a conditional variational",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 280,
+ 505,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 505,
+ 293
+ ],
+ "score": 1.0,
+ "content": "autoencoder (CVAE) to handle residual uncertainty at these bottlenecks. In typical fixed-time CVAE",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 291,
+ 505,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 209,
+ 304
+ ],
+ "score": 1.0,
+ "content": "predictors targeting time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 209,
+ 294,
+ 216,
+ 302
+ ],
+ "score": 0.76,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 216,
+ 291,
+ 327,
+ 304
+ ],
+ "score": 1.0,
+ "content": ", variations in a latent code",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 328,
+ 293,
+ 334,
+ 301
+ ],
+ "score": 0.77,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 335,
+ 291,
+ 418,
+ 304
+ ],
+ "score": 1.0,
+ "content": "input to a generator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 419,
+ 293,
+ 448,
+ 303
+ ],
+ "score": 0.92,
+ "content": "G ( c , z )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 449,
+ 291,
+ 505,
+ 304
+ ],
+ "score": 1.0,
+ "content": "must capture",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 104,
+ 302,
+ 506,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 302,
+ 172,
+ 315
+ ],
+ "score": 1.0,
+ "content": "stochasticity in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 172,
+ 304,
+ 184,
+ 313
+ ],
+ "score": 0.86,
+ "content": "x _ { \\tau }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 185,
+ 302,
+ 266,
+ 315
+ ],
+ "score": 1.0,
+ "content": ". At training time,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 266,
+ 304,
+ 273,
+ 312
+ ],
+ "score": 0.73,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 273,
+ 302,
+ 444,
+ 315
+ ],
+ "score": 1.0,
+ "content": "is sampled from a posterior distribution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 445,
+ 303,
+ 482,
+ 315
+ ],
+ "score": 0.89,
+ "content": "q _ { \\phi } ( z | x _ { \\tau } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 482,
+ 302,
+ 506,
+ 315
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 313,
+ 505,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 313,
+ 154,
+ 326
+ ],
+ "score": 1.0,
+ "content": "parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 154,
+ 314,
+ 161,
+ 325
+ ],
+ "score": 0.85,
+ "content": "\\phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 162,
+ 313,
+ 356,
+ 326
+ ],
+ "score": 1.0,
+ "content": ", represented by a neural network. At test time,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 356,
+ 315,
+ 362,
+ 323
+ ],
+ "score": 0.79,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 363,
+ 313,
+ 462,
+ 326
+ ],
+ "score": 1.0,
+ "content": "is sampled from a prior",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 463,
+ 315,
+ 481,
+ 325
+ ],
+ "score": 0.89,
+ "content": "p ( z )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 482,
+ 313,
+ 505,
+ 326
+ ],
+ "score": 1.0,
+ "content": ". The",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 324,
+ 436,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 324,
+ 436,
+ 338
+ ],
+ "score": 1.0,
+ "content": "training loss combines a log-likelihood term with a KL-divergence from the prior:",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 14.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 170,
+ 343,
+ 440,
+ 359
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 170,
+ 343,
+ 440,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 170,
+ 343,
+ 440,
+ 359
+ ],
+ "score": 0.88,
+ "content": "\\begin{array} { r } { \\mathcal { L } _ { \\mathrm { c v a e } } ( G , \\phi ) = D _ { K L } ( q _ { \\phi } ( z | x _ { \\tau } ) , p ( z ) ) - \\mathbb { E } _ { z \\sim q _ { \\phi } ( z | x _ { \\tau } ) } \\ln p _ { G } ( x _ { \\tau } | c , z ) , } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "b6d76c5d7212d1973d6ab06b4a20053e8b9be5d284c7da59a31de2a21fdaba9c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 18,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 170,
+ 343,
+ 440,
+ 359
+ ],
+ "spans": [],
+ "index": 18
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 365,
+ 506,
+ 443
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 366,
+ 506,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 366,
+ 193,
+ 378
+ ],
+ "score": 1.0,
+ "content": "where we might set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 368,
+ 207,
+ 377
+ ],
+ "score": 0.85,
+ "content": "p _ { G }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 207,
+ 366,
+ 492,
+ 378
+ ],
+ "score": 1.0,
+ "content": "to a Laplacian distribution such that the second term reduces to a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 493,
+ 366,
+ 502,
+ 376
+ ],
+ "score": 0.54,
+ "content": "l 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 502,
+ 366,
+ 506,
+ 378
+ ],
+ "score": 1.0,
+ "content": "-",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 376,
+ 505,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 377,
+ 187,
+ 389
+ ],
+ "score": 1.0,
+ "content": "reconstruction loss",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 187,
+ 376,
+ 339,
+ 389
+ ],
+ "score": 0.89,
+ "content": "- \\ln p _ { G } ( x _ { \\tau } | \\bar { c , { z } } ) = \\| \\boldsymbol { G } ( c , { z } ) - \\boldsymbol { x } _ { \\tau } \\| _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 339,
+ 377,
+ 505,
+ 389
+ ],
+ "score": 1.0,
+ "content": ". In a time-agnostic CVAE, rather than",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 387,
+ 506,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 387,
+ 257,
+ 401
+ ],
+ "score": 1.0,
+ "content": "capturing stochasticity at a fixed time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 258,
+ 390,
+ 264,
+ 398
+ ],
+ "score": 0.68,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 264,
+ 387,
+ 268,
+ 401
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 268,
+ 390,
+ 275,
+ 398
+ ],
+ "score": 0.65,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 387,
+ 506,
+ 401
+ ],
+ "score": 1.0,
+ "content": "must now capture stochasticity at bottlenecks: e.g., when",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 398,
+ 505,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 398,
+ 312,
+ 411
+ ],
+ "score": 1.0,
+ "content": "the agent crosses one of the asterisks in the maze of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 313,
+ 399,
+ 334,
+ 410
+ ],
+ "score": 0.26,
+ "content": "\\mathrm { F i g ~ 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 334,
+ 398,
+ 505,
+ 411
+ ],
+ "score": 1.0,
+ "content": ", which pose is it in? The bottleneck’s time",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 409,
+ 505,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 409,
+ 505,
+ 423
+ ],
+ "score": 1.0,
+ "content": "index varies and is not known in advance. For computational reasons (see Appendix B), we pass",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 420,
+ 505,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 171,
+ 433
+ ],
+ "score": 1.0,
+ "content": "the entire video",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 171,
+ 421,
+ 181,
+ 430
+ ],
+ "score": 0.83,
+ "content": "X",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 420,
+ 289,
+ 433
+ ],
+ "score": 1.0,
+ "content": "into the inference network",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 289,
+ 422,
+ 300,
+ 433
+ ],
+ "score": 0.86,
+ "content": "q _ { \\phi }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 300,
+ 420,
+ 505,
+ 433
+ ],
+ "score": 1.0,
+ "content": ", similar to Babaeizadeh et al. (2018). The negative",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 432,
+ 344,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 432,
+ 344,
+ 444
+ ],
+ "score": 1.0,
+ "content": "log-likelihood term is adapted to be a minimum-over-time:",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 155,
+ 451,
+ 455,
+ 470
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 155,
+ 451,
+ 455,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 155,
+ 451,
+ 455,
+ 470
+ ],
+ "score": 0.91,
+ "content": "\\mathcal { L } _ { \\mathrm { c v a e } } ( G , \\phi ) = D _ { K L } ( q _ { \\phi } ( z | X ) , p ( z ) ) + \\operatorname* { m i n } _ { t \\in \\mathbb { T } } \\mathbb { E } _ { z \\sim q _ { \\phi } ( z | X ) } \\left[ - \\ln p _ { G } ( x _ { t } | c , z ) \\right] .",
+ "type": "interline_equation",
+ "image_path": "e2f011e4146b6cc813ecc2a11f1d6b637c34936b5589f7d57964fdc9b232252f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 26,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 155,
+ 451,
+ 455,
+ 470
+ ],
+ "spans": [],
+ "index": 26
+ }
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 487,
+ 396,
+ 500
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 487,
+ 397,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 487,
+ 397,
+ 501
+ ],
+ "score": 1.0,
+ "content": "3.5 COMBINED LOSS, NETWORK ARCHITECTURE, AND TRAINING",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 507,
+ 504,
+ 531
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 507,
+ 505,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 507,
+ 505,
+ 520
+ ],
+ "score": 1.0,
+ "content": "We train time-agnostic CVAE-GANs with the following combination of a generalized minimum loss",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 519,
+ 325,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 519,
+ 325,
+ 532
+ ],
+ "score": 1.0,
+ "content": "(Sec 3.2) and the CVAE KL divergence loss (Sec 3.4):",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 175,
+ 536,
+ 435,
+ 597
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 175,
+ 536,
+ 435,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 175,
+ 536,
+ 435,
+ 597
+ ],
+ "score": 0.94,
+ "content": "\\begin{array} { r l } & { G ^ { * } = \\underset { G } { \\arg \\operatorname* { m i n } } \\underset { \\phi } { \\operatorname* { m i n } } \\left[ D _ { K L } ( q _ { \\phi } ( \\boldsymbol { z } | \\boldsymbol { X } ) , p ( \\boldsymbol { z } ) ) + \\mathcal { E } _ { \\mathrm { a r g } \\operatorname* { m i n } _ { t \\in \\mathbb { T } } } \\varepsilon _ { t } ^ { \\prime } \\right] , } \\\\ & { \\mathcal { E } _ { t } = \\underset { D , D ^ { \\prime } } { \\operatorname* { m a x } } \\mathcal { L } _ { \\mathrm { c g a n } } ^ { t } ( G , D _ { t } ) + \\mathcal { L } _ { \\mathrm { c v a e - g a n } } ^ { t } ( G , D _ { t } ^ { \\prime } ) + \\| G ( \\boldsymbol { c } , \\boldsymbol { z } ) - x _ { t } \\| _ { 1 } , } \\\\ & { \\mathcal { E } _ { t } ^ { \\prime } = \\| G ( \\boldsymbol { c } , \\boldsymbol { z } ) - x _ { t } \\| _ { 1 } / w ( t ) . } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "cce1881ed319695652dcf6e898c2cb7e7029038e0b7b830af098a8f74e1e396c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 31,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 175,
+ 536,
+ 435,
+ 556.3333333333334
+ ],
+ "spans": [],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 175,
+ 556.3333333333334,
+ 435,
+ 576.6666666666667
+ ],
+ "spans": [],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 175,
+ 576.6666666666667,
+ 435,
+ 597.0000000000001
+ ],
+ "spans": [],
+ "index": 32
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 602,
+ 505,
+ 695
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 603,
+ 506,
+ 615
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 603,
+ 171,
+ 615
+ ],
+ "score": 1.0,
+ "content": "The outer error",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 171,
+ 604,
+ 181,
+ 614
+ ],
+ "score": 0.87,
+ "content": "\\mathcal { E } _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 603,
+ 476,
+ 615
+ ],
+ "score": 1.0,
+ "content": "absorbs the CGAN discriminator errors (Sec 3.3), while the inner error",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 477,
+ 603,
+ 487,
+ 615
+ ],
+ "score": 0.88,
+ "content": "\\mathcal { E } _ { t } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 603,
+ 506,
+ 615
+ ],
+ "score": 1.0,
+ "content": "is a",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 614,
+ 505,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 614,
+ 136,
+ 626
+ ],
+ "score": 1.0,
+ "content": "simple",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 136,
+ 614,
+ 146,
+ 625
+ ],
+ "score": 0.87,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 146,
+ 614,
+ 354,
+ 626
+ ],
+ "score": 1.0,
+ "content": "error, weighted by user-specified time preferences",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 354,
+ 614,
+ 374,
+ 626
+ ],
+ "score": 0.91,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 374,
+ 614,
+ 505,
+ 626
+ ],
+ "score": 1.0,
+ "content": "(Sec 3.2). Omitting GAN error",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 624,
+ 505,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 624,
+ 142,
+ 638
+ ],
+ "score": 1.0,
+ "content": "terms in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 142,
+ 625,
+ 153,
+ 637
+ ],
+ "score": 0.89,
+ "content": "\\mathcal { E } _ { t } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 153,
+ 624,
+ 505,
+ 638
+ ],
+ "score": 1.0,
+ "content": "helps stabilize training, since learned errors may not always be meaningful especially",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 101,
+ 636,
+ 505,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 101,
+ 641,
+ 144,
+ 679
+ ],
+ "score": 1.0,
+ "content": "includesthe prior",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 106,
+ 636,
+ 505,
+ 649
+ ],
+ "score": 1.0,
+ "content": "early on in training. As in VAE-GANs (Larsen et al., 2016; Lee et al., 2018), the training objective",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 163,
+ 641,
+ 191,
+ 679
+ ],
+ "score": 1.0,
+ "content": "w term , while",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 193,
+ 647,
+ 226,
+ 659
+ ],
+ "score": 0.92,
+ "content": "\\mathcal { L } _ { \\mathrm { c v a e - g a n } } ^ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 226,
+ 641,
+ 286,
+ 679
+ ],
+ "score": 1.0,
+ "content": ", analogous to instead sample",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 286,
+ 647,
+ 308,
+ 660
+ ],
+ "score": 0.9,
+ "content": "\\mathcal { L } _ { \\mathrm { c g a n } } ^ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 308,
+ 641,
+ 376,
+ 679
+ ],
+ "score": 1.0,
+ "content": "(Eq 7). We set om the posterior",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 388,
+ 647,
+ 409,
+ 659
+ ],
+ "score": 0.91,
+ "content": "\\mathcal { L } _ { \\mathrm { c g a n } } ^ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 412,
+ 641,
+ 474,
+ 679
+ ],
+ "score": 1.0,
+ "content": "to use samples , and employs s",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 474,
+ 649,
+ 481,
+ 657
+ ],
+ "score": 0.75,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 482,
+ 641,
+ 505,
+ 679
+ ],
+ "score": 1.0,
+ "content": "fromarate",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 144,
+ 659,
+ 411,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 144,
+ 659,
+ 162,
+ 672
+ ],
+ "score": 0.92,
+ "content": "p ( z )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 192,
+ 660,
+ 225,
+ 672
+ ],
+ "score": 0.91,
+ "content": "\\mathcal { L } _ { \\mathrm { c v a e - g a n } } ^ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 293,
+ 662,
+ 299,
+ 669
+ ],
+ "score": 0.78,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 376,
+ 660,
+ 411,
+ 672
+ ],
+ "score": 0.93,
+ "content": "q _ { \\phi } ( z | X )",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 505,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 167,
+ 685
+ ],
+ "score": 1.0,
+ "content": "discriminators",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 167,
+ 672,
+ 189,
+ 684
+ ],
+ "score": 0.92,
+ "content": "\\{ D _ { t } ^ { \\prime } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 190,
+ 671,
+ 212,
+ 685
+ ],
+ "score": 1.0,
+ "content": ". The",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 212,
+ 673,
+ 222,
+ 683
+ ],
+ "score": 0.86,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 222,
+ 671,
+ 295,
+ 685
+ ],
+ "score": 1.0,
+ "content": "loss also samples",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 295,
+ 675,
+ 302,
+ 682
+ ],
+ "score": 0.76,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 302,
+ 671,
+ 505,
+ 685
+ ],
+ "score": 1.0,
+ "content": "from the posterior. We omit expectations over the",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 682,
+ 258,
+ 696
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 682,
+ 152,
+ 696
+ ],
+ "score": 1.0,
+ "content": "VAE latent",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 153,
+ 685,
+ 159,
+ 693
+ ],
+ "score": 0.76,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 159,
+ 682,
+ 258,
+ 696
+ ],
+ "score": 1.0,
+ "content": "to keep notation simple.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 700,
+ 505,
+ 734
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "Frame generation in the predictor involves first generating appearance flow-transformed input",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "frames (Zhou et al., 2016) and a frame with new uncopied pixels. These frames are masked and",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 722,
+ 460,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 722,
+ 460,
+ 734
+ ],
+ "score": 1.0,
+ "content": "averaged to produce the output. Full architecture and training details are in Appendix C.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "page_idx": 4,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 26,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "score": 1.0,
+ "content": "5",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 81,
+ 505,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 506,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 295,
+ 96
+ ],
+ "score": 1.0,
+ "content": "To make this compatible with TAP, we train",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 296,
+ 83,
+ 310,
+ 95
+ ],
+ "score": 0.86,
+ "content": "| \\mathrm { T } |",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 310,
+ 81,
+ 374,
+ 96
+ ],
+ "score": 1.0,
+ "content": "discriminators",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 374,
+ 82,
+ 397,
+ 95
+ ],
+ "score": 0.92,
+ "content": "\\{ D _ { t } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 397,
+ 81,
+ 506,
+ 96
+ ],
+ "score": 1.0,
+ "content": ", one per timestep. Then,",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 359,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 159,
+ 106
+ ],
+ "score": 1.0,
+ "content": "analogous to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 159,
+ 94,
+ 180,
+ 105
+ ],
+ "score": 0.28,
+ "content": "\\operatorname { E q } 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 180,
+ 93,
+ 359,
+ 106
+ ],
+ "score": 1.0,
+ "content": ", we may define a time-agnostic CGAN loss:",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5,
+ "bbox_fs": [
+ 105,
+ 81,
+ 506,
+ 106
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 141,
+ 111,
+ 469,
+ 163
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 111,
+ 469,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 111,
+ 469,
+ 163
+ ],
+ "score": 0.93,
+ "content": "\\begin{array} { r l } & { G ^ { * } = \\underset { G } { \\arg \\operatorname* { m i n } } \\underset { t \\in T } { \\operatorname* { m i n } } \\underset { D _ { t } } { \\operatorname* { m a x } } \\mathcal { L } _ { \\mathrm { c g a n } } ^ { t } ( G , D _ { t } ) , } \\\\ & { \\mathcal { L } _ { \\mathrm { c g a n } } ^ { t } ( G , D _ { t } ) = \\log D _ { t } ( c , x _ { t } ) + \\log ( 1 - D _ { t } ( c , G ( c ) ) ) + \\underset { t ^ { \\prime } \\neq t } { \\sum } \\log ( 1 - D _ { t } ( c , x _ { t ^ { \\prime } } ) ) , } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "642032ec3b2c5e42befe78ad2d47ede55bef3f7c187593e8ad908907a0646f0a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 141,
+ 111,
+ 469,
+ 128.33333333333334
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 141,
+ 128.33333333333334,
+ 469,
+ 145.66666666666669
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 141,
+ 145.66666666666669,
+ 469,
+ 163.00000000000003
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 169,
+ 506,
+ 236
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 169,
+ 505,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 169,
+ 127,
+ 182
+ ],
+ "score": 1.0,
+ "content": "Like",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 128,
+ 170,
+ 173,
+ 182
+ ],
+ "score": 0.42,
+ "content": "\\mathrm { E q } 6 , \\mathrm { E q } 7",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 173,
+ 169,
+ 441,
+ 182
+ ],
+ "score": 1.0,
+ "content": "defines a cross-entropy loss. The first two terms are analogous to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 441,
+ 170,
+ 462,
+ 182
+ ],
+ "score": 0.37,
+ "content": "\\operatorname { E q } 6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 463,
+ 169,
+ 505,
+ 182
+ ],
+ "score": 1.0,
+ "content": "— for the",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 107,
+ 180,
+ 505,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 182,
+ 111,
+ 191
+ ],
+ "score": 0.75,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 112,
+ 180,
+ 198,
+ 194
+ ],
+ "score": 1.0,
+ "content": "-th discriminator, the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 198,
+ 182,
+ 203,
+ 191
+ ],
+ "score": 0.76,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 180,
+ 505,
+ 194
+ ],
+ "score": 1.0,
+ "content": "-th frame provides a positive, and the generated frame provides a negative",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 191,
+ 507,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 191,
+ 430,
+ 205
+ ],
+ "score": 1.0,
+ "content": "instance. The third term treats ground truth video frames occurring at other times",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 431,
+ 193,
+ 453,
+ 203
+ ],
+ "score": 0.85,
+ "content": "x _ { t ^ { \\prime } \\neq t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 453,
+ 191,
+ 507,
+ 205
+ ],
+ "score": 1.0,
+ "content": "as negatives.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 202,
+ 506,
+ 215
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 202,
+ 332,
+ 215
+ ],
+ "score": 1.0,
+ "content": "In practice, we train a single discriminator network with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 332,
+ 203,
+ 346,
+ 215
+ ],
+ "score": 0.9,
+ "content": "| T |",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 346,
+ 202,
+ 430,
+ 215
+ ],
+ "score": 1.0,
+ "content": "outputs that serve as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 430,
+ 204,
+ 453,
+ 215
+ ],
+ "score": 0.9,
+ "content": "\\{ D _ { t } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 453,
+ 202,
+ 506,
+ 215
+ ],
+ "score": 1.0,
+ "content": ". Further, for",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 214,
+ 505,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 214,
+ 358,
+ 226
+ ],
+ "score": 1.0,
+ "content": "computational efficiency, we approximate the summation over",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 359,
+ 214,
+ 383,
+ 225
+ ],
+ "score": 0.91,
+ "content": "t ^ { \\prime } \\neq t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 383,
+ 214,
+ 505,
+ 226
+ ],
+ "score": 1.0,
+ "content": "by sampling a single frame at",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 225,
+ 463,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 225,
+ 463,
+ 237
+ ],
+ "score": 1.0,
+ "content": "random for each training video at each iteration. Appendix A provides additional details.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 7.5,
+ "bbox_fs": [
+ 105,
+ 169,
+ 507,
+ 237
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 249,
+ 296,
+ 261
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 249,
+ 297,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 249,
+ 297,
+ 261
+ ],
+ "score": 1.0,
+ "content": "3.4 TIME-AGNOSTIC CONDITIONAL VAES",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 269,
+ 505,
+ 336
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 270,
+ 505,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 270,
+ 505,
+ 281
+ ],
+ "score": 1.0,
+ "content": "While TAP targets low-uncertainty bottleneck states, it may be integrated with a conditional variational",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 280,
+ 505,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 505,
+ 293
+ ],
+ "score": 1.0,
+ "content": "autoencoder (CVAE) to handle residual uncertainty at these bottlenecks. In typical fixed-time CVAE",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 291,
+ 505,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 209,
+ 304
+ ],
+ "score": 1.0,
+ "content": "predictors targeting time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 209,
+ 294,
+ 216,
+ 302
+ ],
+ "score": 0.76,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 216,
+ 291,
+ 327,
+ 304
+ ],
+ "score": 1.0,
+ "content": ", variations in a latent code",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 328,
+ 293,
+ 334,
+ 301
+ ],
+ "score": 0.77,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 335,
+ 291,
+ 418,
+ 304
+ ],
+ "score": 1.0,
+ "content": "input to a generator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 419,
+ 293,
+ 448,
+ 303
+ ],
+ "score": 0.92,
+ "content": "G ( c , z )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 449,
+ 291,
+ 505,
+ 304
+ ],
+ "score": 1.0,
+ "content": "must capture",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 104,
+ 302,
+ 506,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 302,
+ 172,
+ 315
+ ],
+ "score": 1.0,
+ "content": "stochasticity in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 172,
+ 304,
+ 184,
+ 313
+ ],
+ "score": 0.86,
+ "content": "x _ { \\tau }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 185,
+ 302,
+ 266,
+ 315
+ ],
+ "score": 1.0,
+ "content": ". At training time,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 266,
+ 304,
+ 273,
+ 312
+ ],
+ "score": 0.73,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 273,
+ 302,
+ 444,
+ 315
+ ],
+ "score": 1.0,
+ "content": "is sampled from a posterior distribution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 445,
+ 303,
+ 482,
+ 315
+ ],
+ "score": 0.89,
+ "content": "q _ { \\phi } ( z | x _ { \\tau } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 482,
+ 302,
+ 506,
+ 315
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 313,
+ 505,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 313,
+ 154,
+ 326
+ ],
+ "score": 1.0,
+ "content": "parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 154,
+ 314,
+ 161,
+ 325
+ ],
+ "score": 0.85,
+ "content": "\\phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 162,
+ 313,
+ 356,
+ 326
+ ],
+ "score": 1.0,
+ "content": ", represented by a neural network. At test time,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 356,
+ 315,
+ 362,
+ 323
+ ],
+ "score": 0.79,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 363,
+ 313,
+ 462,
+ 326
+ ],
+ "score": 1.0,
+ "content": "is sampled from a prior",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 463,
+ 315,
+ 481,
+ 325
+ ],
+ "score": 0.89,
+ "content": "p ( z )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 482,
+ 313,
+ 505,
+ 326
+ ],
+ "score": 1.0,
+ "content": ". The",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 324,
+ 436,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 324,
+ 436,
+ 338
+ ],
+ "score": 1.0,
+ "content": "training loss combines a log-likelihood term with a KL-divergence from the prior:",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 14.5,
+ "bbox_fs": [
+ 104,
+ 270,
+ 506,
+ 338
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 170,
+ 343,
+ 440,
+ 359
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 170,
+ 343,
+ 440,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 170,
+ 343,
+ 440,
+ 359
+ ],
+ "score": 0.88,
+ "content": "\\begin{array} { r } { \\mathcal { L } _ { \\mathrm { c v a e } } ( G , \\phi ) = D _ { K L } ( q _ { \\phi } ( z | x _ { \\tau } ) , p ( z ) ) - \\mathbb { E } _ { z \\sim q _ { \\phi } ( z | x _ { \\tau } ) } \\ln p _ { G } ( x _ { \\tau } | c , z ) , } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "b6d76c5d7212d1973d6ab06b4a20053e8b9be5d284c7da59a31de2a21fdaba9c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 18,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 170,
+ 343,
+ 440,
+ 359
+ ],
+ "spans": [],
+ "index": 18
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 365,
+ 506,
+ 443
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 366,
+ 506,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 366,
+ 193,
+ 378
+ ],
+ "score": 1.0,
+ "content": "where we might set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 368,
+ 207,
+ 377
+ ],
+ "score": 0.85,
+ "content": "p _ { G }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 207,
+ 366,
+ 492,
+ 378
+ ],
+ "score": 1.0,
+ "content": "to a Laplacian distribution such that the second term reduces to a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 493,
+ 366,
+ 502,
+ 376
+ ],
+ "score": 0.54,
+ "content": "l 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 502,
+ 366,
+ 506,
+ 378
+ ],
+ "score": 1.0,
+ "content": "-",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 376,
+ 505,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 377,
+ 187,
+ 389
+ ],
+ "score": 1.0,
+ "content": "reconstruction loss",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 187,
+ 376,
+ 339,
+ 389
+ ],
+ "score": 0.89,
+ "content": "- \\ln p _ { G } ( x _ { \\tau } | \\bar { c , { z } } ) = \\| \\boldsymbol { G } ( c , { z } ) - \\boldsymbol { x } _ { \\tau } \\| _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 339,
+ 377,
+ 505,
+ 389
+ ],
+ "score": 1.0,
+ "content": ". In a time-agnostic CVAE, rather than",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 387,
+ 506,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 387,
+ 257,
+ 401
+ ],
+ "score": 1.0,
+ "content": "capturing stochasticity at a fixed time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 258,
+ 390,
+ 264,
+ 398
+ ],
+ "score": 0.68,
+ "content": "\\tau",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 264,
+ 387,
+ 268,
+ 401
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 268,
+ 390,
+ 275,
+ 398
+ ],
+ "score": 0.65,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 387,
+ 506,
+ 401
+ ],
+ "score": 1.0,
+ "content": "must now capture stochasticity at bottlenecks: e.g., when",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 398,
+ 505,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 398,
+ 312,
+ 411
+ ],
+ "score": 1.0,
+ "content": "the agent crosses one of the asterisks in the maze of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 313,
+ 399,
+ 334,
+ 410
+ ],
+ "score": 0.26,
+ "content": "\\mathrm { F i g ~ 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 334,
+ 398,
+ 505,
+ 411
+ ],
+ "score": 1.0,
+ "content": ", which pose is it in? The bottleneck’s time",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 409,
+ 505,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 409,
+ 505,
+ 423
+ ],
+ "score": 1.0,
+ "content": "index varies and is not known in advance. For computational reasons (see Appendix B), we pass",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 420,
+ 505,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 171,
+ 433
+ ],
+ "score": 1.0,
+ "content": "the entire video",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 171,
+ 421,
+ 181,
+ 430
+ ],
+ "score": 0.83,
+ "content": "X",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 420,
+ 289,
+ 433
+ ],
+ "score": 1.0,
+ "content": "into the inference network",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 289,
+ 422,
+ 300,
+ 433
+ ],
+ "score": 0.86,
+ "content": "q _ { \\phi }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 300,
+ 420,
+ 505,
+ 433
+ ],
+ "score": 1.0,
+ "content": ", similar to Babaeizadeh et al. (2018). The negative",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 432,
+ 344,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 432,
+ 344,
+ 444
+ ],
+ "score": 1.0,
+ "content": "log-likelihood term is adapted to be a minimum-over-time:",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22,
+ "bbox_fs": [
+ 105,
+ 366,
+ 506,
+ 444
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 155,
+ 451,
+ 455,
+ 470
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 155,
+ 451,
+ 455,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 155,
+ 451,
+ 455,
+ 470
+ ],
+ "score": 0.91,
+ "content": "\\mathcal { L } _ { \\mathrm { c v a e } } ( G , \\phi ) = D _ { K L } ( q _ { \\phi } ( z | X ) , p ( z ) ) + \\operatorname* { m i n } _ { t \\in \\mathbb { T } } \\mathbb { E } _ { z \\sim q _ { \\phi } ( z | X ) } \\left[ - \\ln p _ { G } ( x _ { t } | c , z ) \\right] .",
+ "type": "interline_equation",
+ "image_path": "e2f011e4146b6cc813ecc2a11f1d6b637c34936b5589f7d57964fdc9b232252f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 26,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 155,
+ 451,
+ 455,
+ 470
+ ],
+ "spans": [],
+ "index": 26
+ }
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 487,
+ 396,
+ 500
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 487,
+ 397,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 487,
+ 397,
+ 501
+ ],
+ "score": 1.0,
+ "content": "3.5 COMBINED LOSS, NETWORK ARCHITECTURE, AND TRAINING",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 507,
+ 504,
+ 531
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 507,
+ 505,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 507,
+ 505,
+ 520
+ ],
+ "score": 1.0,
+ "content": "We train time-agnostic CVAE-GANs with the following combination of a generalized minimum loss",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 519,
+ 325,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 519,
+ 325,
+ 532
+ ],
+ "score": 1.0,
+ "content": "(Sec 3.2) and the CVAE KL divergence loss (Sec 3.4):",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28.5,
+ "bbox_fs": [
+ 106,
+ 507,
+ 505,
+ 532
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 175,
+ 536,
+ 435,
+ 597
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 175,
+ 536,
+ 435,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 175,
+ 536,
+ 435,
+ 597
+ ],
+ "score": 0.94,
+ "content": "\\begin{array} { r l } & { G ^ { * } = \\underset { G } { \\arg \\operatorname* { m i n } } \\underset { \\phi } { \\operatorname* { m i n } } \\left[ D _ { K L } ( q _ { \\phi } ( \\boldsymbol { z } | \\boldsymbol { X } ) , p ( \\boldsymbol { z } ) ) + \\mathcal { E } _ { \\mathrm { a r g } \\operatorname* { m i n } _ { t \\in \\mathbb { T } } } \\varepsilon _ { t } ^ { \\prime } \\right] , } \\\\ & { \\mathcal { E } _ { t } = \\underset { D , D ^ { \\prime } } { \\operatorname* { m a x } } \\mathcal { L } _ { \\mathrm { c g a n } } ^ { t } ( G , D _ { t } ) + \\mathcal { L } _ { \\mathrm { c v a e - g a n } } ^ { t } ( G , D _ { t } ^ { \\prime } ) + \\| G ( \\boldsymbol { c } , \\boldsymbol { z } ) - x _ { t } \\| _ { 1 } , } \\\\ & { \\mathcal { E } _ { t } ^ { \\prime } = \\| G ( \\boldsymbol { c } , \\boldsymbol { z } ) - x _ { t } \\| _ { 1 } / w ( t ) . } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "cce1881ed319695652dcf6e898c2cb7e7029038e0b7b830af098a8f74e1e396c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 31,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 175,
+ 536,
+ 435,
+ 556.3333333333334
+ ],
+ "spans": [],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 175,
+ 556.3333333333334,
+ 435,
+ 576.6666666666667
+ ],
+ "spans": [],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 175,
+ 576.6666666666667,
+ 435,
+ 597.0000000000001
+ ],
+ "spans": [],
+ "index": 32
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 602,
+ 505,
+ 695
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 603,
+ 506,
+ 615
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 603,
+ 171,
+ 615
+ ],
+ "score": 1.0,
+ "content": "The outer error",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 171,
+ 604,
+ 181,
+ 614
+ ],
+ "score": 0.87,
+ "content": "\\mathcal { E } _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 181,
+ 603,
+ 476,
+ 615
+ ],
+ "score": 1.0,
+ "content": "absorbs the CGAN discriminator errors (Sec 3.3), while the inner error",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 477,
+ 603,
+ 487,
+ 615
+ ],
+ "score": 0.88,
+ "content": "\\mathcal { E } _ { t } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 603,
+ 506,
+ 615
+ ],
+ "score": 1.0,
+ "content": "is a",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 614,
+ 505,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 614,
+ 136,
+ 626
+ ],
+ "score": 1.0,
+ "content": "simple",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 136,
+ 614,
+ 146,
+ 625
+ ],
+ "score": 0.87,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 146,
+ 614,
+ 354,
+ 626
+ ],
+ "score": 1.0,
+ "content": "error, weighted by user-specified time preferences",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 354,
+ 614,
+ 374,
+ 626
+ ],
+ "score": 0.91,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 374,
+ 614,
+ 505,
+ 626
+ ],
+ "score": 1.0,
+ "content": "(Sec 3.2). Omitting GAN error",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 624,
+ 505,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 624,
+ 142,
+ 638
+ ],
+ "score": 1.0,
+ "content": "terms in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 142,
+ 625,
+ 153,
+ 637
+ ],
+ "score": 0.89,
+ "content": "\\mathcal { E } _ { t } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 153,
+ 624,
+ 505,
+ 638
+ ],
+ "score": 1.0,
+ "content": "helps stabilize training, since learned errors may not always be meaningful especially",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 101,
+ 636,
+ 505,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 101,
+ 641,
+ 144,
+ 679
+ ],
+ "score": 1.0,
+ "content": "includesthe prior",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 106,
+ 636,
+ 505,
+ 649
+ ],
+ "score": 1.0,
+ "content": "early on in training. As in VAE-GANs (Larsen et al., 2016; Lee et al., 2018), the training objective",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 163,
+ 641,
+ 191,
+ 679
+ ],
+ "score": 1.0,
+ "content": "w term , while",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 193,
+ 647,
+ 226,
+ 659
+ ],
+ "score": 0.92,
+ "content": "\\mathcal { L } _ { \\mathrm { c v a e - g a n } } ^ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 226,
+ 641,
+ 286,
+ 679
+ ],
+ "score": 1.0,
+ "content": ", analogous to instead sample",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 286,
+ 647,
+ 308,
+ 660
+ ],
+ "score": 0.9,
+ "content": "\\mathcal { L } _ { \\mathrm { c g a n } } ^ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 308,
+ 641,
+ 376,
+ 679
+ ],
+ "score": 1.0,
+ "content": "(Eq 7). We set om the posterior",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 388,
+ 647,
+ 409,
+ 659
+ ],
+ "score": 0.91,
+ "content": "\\mathcal { L } _ { \\mathrm { c g a n } } ^ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 412,
+ 641,
+ 474,
+ 679
+ ],
+ "score": 1.0,
+ "content": "to use samples , and employs s",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 474,
+ 649,
+ 481,
+ 657
+ ],
+ "score": 0.75,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 482,
+ 641,
+ 505,
+ 679
+ ],
+ "score": 1.0,
+ "content": "fromarate",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 144,
+ 659,
+ 411,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 144,
+ 659,
+ 162,
+ 672
+ ],
+ "score": 0.92,
+ "content": "p ( z )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 192,
+ 660,
+ 225,
+ 672
+ ],
+ "score": 0.91,
+ "content": "\\mathcal { L } _ { \\mathrm { c v a e - g a n } } ^ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 293,
+ 662,
+ 299,
+ 669
+ ],
+ "score": 0.78,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 376,
+ 660,
+ 411,
+ 672
+ ],
+ "score": 0.93,
+ "content": "q _ { \\phi } ( z | X )",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 505,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 167,
+ 685
+ ],
+ "score": 1.0,
+ "content": "discriminators",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 167,
+ 672,
+ 189,
+ 684
+ ],
+ "score": 0.92,
+ "content": "\\{ D _ { t } ^ { \\prime } \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 190,
+ 671,
+ 212,
+ 685
+ ],
+ "score": 1.0,
+ "content": ". The",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 212,
+ 673,
+ 222,
+ 683
+ ],
+ "score": 0.86,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 222,
+ 671,
+ 295,
+ 685
+ ],
+ "score": 1.0,
+ "content": "loss also samples",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 295,
+ 675,
+ 302,
+ 682
+ ],
+ "score": 0.76,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 302,
+ 671,
+ 505,
+ 685
+ ],
+ "score": 1.0,
+ "content": "from the posterior. We omit expectations over the",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 682,
+ 258,
+ 696
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 682,
+ 152,
+ 696
+ ],
+ "score": 1.0,
+ "content": "VAE latent",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 153,
+ 685,
+ 159,
+ 693
+ ],
+ "score": 0.76,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 159,
+ 682,
+ 258,
+ 696
+ ],
+ "score": 1.0,
+ "content": "to keep notation simple.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 36,
+ "bbox_fs": [
+ 101,
+ 603,
+ 506,
+ 696
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 700,
+ 505,
+ 734
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "Frame generation in the predictor involves first generating appearance flow-transformed input",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "frames (Zhou et al., 2016) and a frame with new uncopied pixels. These frames are masked and",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 722,
+ 460,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 722,
+ 460,
+ 734
+ ],
+ "score": 1.0,
+ "content": "averaged to produce the output. Full architecture and training details are in Appendix C.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 41,
+ "bbox_fs": [
+ 105,
+ 700,
+ 506,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 106,
+ 72,
+ 378,
+ 182
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 106,
+ 72,
+ 378,
+ 182
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 378,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 378,
+ 182
+ ],
+ "score": 0.968,
+ "type": "image",
+ "image_path": "5770695687526489da2068203ed0e833315b117d7b1003431ea451379f1cd9f3.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 378,
+ 108.66666666666666
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 108.66666666666666,
+ 378,
+ 145.33333333333331
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 145.33333333333331,
+ 378,
+ 181.99999999999997
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 191,
+ 378,
+ 211
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 190,
+ 379,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 190,
+ 379,
+ 202
+ ],
+ "score": 1.0,
+ "content": "Figure 3: (Best seen in pdf) One sample episode each for grasping, pick-and-",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 201,
+ 286,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 201,
+ 286,
+ 212
+ ],
+ "score": 1.0,
+ "content": "place, and pushing. Time overlaid on each frame.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5
+ }
+ ],
+ "index": 3.25
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 390,
+ 85,
+ 498,
+ 159
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 390,
+ 85,
+ 498,
+ 159
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 390,
+ 85,
+ 498,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 390,
+ 85,
+ 498,
+ 159
+ ],
+ "score": 0.943,
+ "type": "image",
+ "image_path": "c276a2df0e0bfd87996b9ced2d4c7e08dda70fc8eb2ef52b070ed396f8e3d0f5.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 390,
+ 85,
+ 498,
+ 122.0
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 390,
+ 122.0,
+ 498,
+ 159.0
+ ],
+ "spans": [],
+ "index": 3
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 387,
+ 160,
+ 505,
+ 210
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 387,
+ 160,
+ 504,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 387,
+ 160,
+ 495,
+ 171
+ ],
+ "score": 1.0,
+ "content": "Figure 4: Forward prediction",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 495,
+ 160,
+ 504,
+ 170
+ ],
+ "score": 0.81,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 387,
+ 170,
+ 505,
+ 181
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 387,
+ 170,
+ 505,
+ 181
+ ],
+ "score": 1.0,
+ "content": "error for grasping. TAP methods",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 387,
+ 180,
+ 506,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 387,
+ 180,
+ 506,
+ 190
+ ],
+ "score": 1.0,
+ "content": "(red) perform better than fixed-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 387,
+ 190,
+ 505,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 387,
+ 190,
+ 505,
+ 201
+ ],
+ "score": 1.0,
+ "content": "time predictors over all time",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 386,
+ 200,
+ 411,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 386,
+ 200,
+ 411,
+ 213
+ ],
+ "score": 1.0,
+ "content": "steps.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 5.75
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 224,
+ 200,
+ 236
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 223,
+ 201,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 223,
+ 201,
+ 238
+ ],
+ "score": 1.0,
+ "content": "4 EXPERIMENTS",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 247,
+ 505,
+ 314
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 248,
+ 505,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 248,
+ 505,
+ 259
+ ],
+ "score": 1.0,
+ "content": "We have proposed a time-agnostic prediction (TAP) paradigm that is different from the fixed-time",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 258,
+ 506,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 506,
+ 272
+ ],
+ "score": 1.0,
+ "content": "paradigm followed in prior prediction work. In our experiments, we focus on comparing TAP against",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 104,
+ 269,
+ 505,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 269,
+ 505,
+ 282
+ ],
+ "score": 1.0,
+ "content": "a representative fixed-time prediction model, keeping network architectures fixed. We use three",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 280,
+ 507,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 325,
+ 294
+ ],
+ "score": 1.0,
+ "content": "simulated robot manipulation settings: object grasping (",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 325,
+ 281,
+ 341,
+ 291
+ ],
+ "score": 0.41,
+ "content": "5 0 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 342,
+ 280,
+ 447,
+ 294
+ ],
+ "score": 1.0,
+ "content": "episodes), pick-and-place",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 447,
+ 281,
+ 464,
+ 291
+ ],
+ "score": 0.44,
+ "content": "7 5 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 464,
+ 280,
+ 507,
+ 294
+ ],
+ "score": 1.0,
+ "content": "episodes),",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 303
+ ],
+ "score": 1.0,
+ "content": "and multi-object pushing (55k episodes). Example episodes from each task are shown in Fig 3 (videos",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 302,
+ 398,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 302,
+ 145,
+ 316
+ ],
+ "score": 1.0,
+ "content": "in Supp).",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 303,
+ 161,
+ 313
+ ],
+ "score": 0.79,
+ "content": "5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 161,
+ 302,
+ 329,
+ 316
+ ],
+ "score": 1.0,
+ "content": "of the data is set aside for testing. We use",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 330,
+ 303,
+ 364,
+ 313
+ ],
+ "score": 0.89,
+ "content": "6 4 \\times 6 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 364,
+ 302,
+ 398,
+ 316
+ ],
+ "score": 1.0,
+ "content": "images.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 319,
+ 505,
+ 374
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 320,
+ 505,
+ 332
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 320,
+ 505,
+ 332
+ ],
+ "score": 1.0,
+ "content": "For grasping (15 frames per episode), the arm moves to a single object on a table, selects a grasp, and",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 329,
+ 506,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 329,
+ 506,
+ 343
+ ],
+ "score": 1.0,
+ "content": "lifts it vertically. For pick-and-place (20 frames), the arm additionally places the lifted object at a new",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 340,
+ 506,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 340,
+ 506,
+ 354
+ ],
+ "score": 1.0,
+ "content": "position before performing a random walk. For pushing (40 frames), two objects are initialized at",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 351,
+ 505,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 505,
+ 366
+ ],
+ "score": 1.0,
+ "content": "random locations and pushed to random final positions. Object shapes and colors in all three settings",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 363,
+ 332,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 363,
+ 332,
+ 376
+ ],
+ "score": 1.0,
+ "content": "are randomly generated. Fig 3 shows example episodes.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 379,
+ 505,
+ 468
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 379,
+ 506,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 379,
+ 506,
+ 392
+ ],
+ "score": 1.0,
+ "content": "Forward prediction. First, we evaluate our approach for forward prediction on grasping. The first",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 389,
+ 506,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 389,
+ 506,
+ 404
+ ],
+ "score": 1.0,
+ "content": "frame (“start”) is provided as input. We train fixed-time baselines (architecture same as ours, using",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 402,
+ 506,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 402,
+ 116,
+ 412
+ ],
+ "score": 0.86,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 117,
+ 402,
+ 506,
+ 414
+ ],
+ "score": 1.0,
+ "content": "and GAN losses same as MIN and GENMIN) that target predictions at exactly 0.25, 0.50, 0.75, 1.0",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 412,
+ 505,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 412,
+ 505,
+ 425
+ ],
+ "score": 1.0,
+ "content": "fraction of the episode length (FIX0.25,. . . , FIX1.00). MIN and GENMIN are TAP with/without the",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 423,
+ 505,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 423,
+ 207,
+ 436
+ ],
+ "score": 1.0,
+ "content": "generalized minimum of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 208,
+ 424,
+ 238,
+ 434
+ ],
+ "score": 0.38,
+ "content": "{ \\tt S e c 3 . 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 239,
+ 423,
+ 505,
+ 436
+ ],
+ "score": 1.0,
+ "content": ". For GENMIN, we evaluate different choices of the time preference",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 433,
+ 504,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 433,
+ 134,
+ 447
+ ],
+ "score": 1.0,
+ "content": "vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 135,
+ 434,
+ 154,
+ 446
+ ],
+ "score": 0.91,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 155,
+ 433,
+ 230,
+ 447
+ ],
+ "score": 1.0,
+ "content": "(Sec 3.2). We set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 231,
+ 434,
+ 303,
+ 446
+ ],
+ "score": 0.92,
+ "content": "w ( t ) = \\beta + t / 1 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 303,
+ 433,
+ 496,
+ 447
+ ],
+ "score": 1.0,
+ "content": ", so that our preference increases linearly from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 496,
+ 435,
+ 504,
+ 446
+ ],
+ "score": 0.82,
+ "content": "\\beta",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 104,
+ 444,
+ 506,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 444,
+ 117,
+ 458
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 117,
+ 446,
+ 142,
+ 456
+ ],
+ "score": 0.89,
+ "content": "\\beta + 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 142,
+ 444,
+ 173,
+ 458
+ ],
+ "score": 1.0,
+ "content": ". Since",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 173,
+ 445,
+ 193,
+ 457
+ ],
+ "score": 0.91,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 194,
+ 444,
+ 313,
+ 458
+ ],
+ "score": 1.0,
+ "content": "applies multiplicatively, low",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 313,
+ 446,
+ 321,
+ 457
+ ],
+ "score": 0.83,
+ "content": "\\beta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 322,
+ 444,
+ 506,
+ 458
+ ],
+ "score": 1.0,
+ "content": "corresponds to high disparity in preferences",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 456,
+ 498,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 457,
+ 109,
+ 469
+ ],
+ "score": 0.0,
+ "content": "",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 110,
+ 457,
+ 140,
+ 468
+ ],
+ "score": 0.89,
+ "content": "\\beta = \\infty",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 140,
+ 457,
+ 428,
+ 469
+ ],
+ "score": 1.0,
+ "content": "reduces to MIN, i.e., no time preference). GENMIN2 is our approach with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 429,
+ 456,
+ 454,
+ 468
+ ],
+ "score": 0.91,
+ "content": "\\beta = 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 455,
+ 457,
+ 498,
+ 469
+ ],
+ "score": 1.0,
+ "content": "and so on.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 27.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 473,
+ 505,
+ 539
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 473,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 473,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "Fig 5 shows example predictions from all methods for the grasping task. In terms of visual quality",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 484,
+ 506,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 484,
+ 506,
+ 497
+ ],
+ "score": 1.0,
+ "content": "of predictions and finding a semantically coherent bottleneck, GENMIN2, GENMIN4, and GENMIN7",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 104,
+ 495,
+ 506,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 495,
+ 506,
+ 508
+ ],
+ "score": 1.0,
+ "content": "perform best—they reliably produce a grasp on the object while it is still on the table. With little or",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 506,
+ 505,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 506,
+ 505,
+ 519
+ ],
+ "score": 1.0,
+ "content": "no time preferences, MIN and GENMIN10 produce images very close to the start, while GENMIN0.5",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 517,
+ 506,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 517,
+ 506,
+ 530
+ ],
+ "score": 1.0,
+ "content": "places too high a value on predictions farther away, and produces blurry images of the object after",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 525,
+ 137,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 525,
+ 137,
+ 542
+ ],
+ "score": 1.0,
+ "content": "lifting.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 34.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 544,
+ 505,
+ 644
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 545,
+ 505,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 545,
+ 408,
+ 556
+ ],
+ "score": 1.0,
+ "content": "Quantitatively, for each method, we report the min and arg min index of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 408,
+ 545,
+ 418,
+ 556
+ ],
+ "score": 0.87,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 418,
+ 545,
+ 505,
+ 556
+ ],
+ "score": 1.0,
+ "content": "distance to all frames",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 505,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 192,
+ 569
+ ],
+ "score": 1.0,
+ "content": "in the video, as “min",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 193,
+ 556,
+ 203,
+ 567
+ ],
+ "score": 0.85,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 555,
+ 505,
+ 569
+ ],
+ "score": 1.0,
+ "content": "err” and “match-step” (“which future ground truth frame is the prediction",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 567,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 567,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": "closest to?”). Fig 4 shows a scatter plot, where each dot or square is one model. TAP (various models",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 578,
+ 506,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 578,
+ 160,
+ 591
+ ],
+ "score": 1.0,
+ "content": "with varying",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 161,
+ 578,
+ 168,
+ 590
+ ],
+ "score": 0.75,
+ "content": "\\beta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 169,
+ 578,
+ 506,
+ 591
+ ],
+ "score": 1.0,
+ "content": ") produces an even larger variation in stepsizes than fixed-time methods explicitly",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 589,
+ 506,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 589,
+ 215,
+ 601
+ ],
+ "score": 1.0,
+ "content": "targeting the entire video",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 216,
+ 589,
+ 248,
+ 599
+ ],
+ "score": 0.48,
+ "content": "\\mathrm { ' } _ { \\mathrm { F I X 0 . 7 5 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 249,
+ 589,
+ 267,
+ 601
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 268,
+ 589,
+ 295,
+ 599
+ ],
+ "score": 0.25,
+ "content": "\\mathrm { F I X } 1 . 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 295,
+ 589,
+ 506,
+ 601
+ ],
+ "score": 1.0,
+ "content": "fall short of producing predictions at 0.75 and 1.0",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 599,
+ 506,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 506,
+ 613
+ ],
+ "score": 1.0,
+ "content": "fraction of the episode length). TAP also produces higher quality predictions (lower error) over that",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 610,
+ 505,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 610,
+ 505,
+ 624
+ ],
+ "score": 1.0,
+ "content": "entire range. From these quantitative and qualitative results, we see that TAP not only successfully",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 621,
+ 505,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 621,
+ 505,
+ 635
+ ],
+ "score": 1.0,
+ "content": "encourages semantically coherent bottleneck predictions, it also produces higher quality predictions",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 633,
+ 326,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 326,
+ 645
+ ],
+ "score": 1.0,
+ "content": "than fixed-time prediction over a range of time offsets.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 42
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 649,
+ 505,
+ 704
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 505,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 505,
+ 662
+ ],
+ "score": 1.0,
+ "content": "Intermediate frame prediction. Next, we evaluate our approaches for bidirectionally conditioned",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 660,
+ 505,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 505,
+ 672
+ ],
+ "score": 1.0,
+ "content": "prediction in all three settings. Initial and final frames are provided as input, and the method is trained",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 505,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 505,
+ 684
+ ],
+ "score": 1.0,
+ "content": "to generate an intermediate frame. The FIX baseline now targets the middle frame. As before, MIN",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 681,
+ 505,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 681,
+ 360,
+ 694
+ ],
+ "score": 1.0,
+ "content": "and GENMIN are our TAP models. The GENMIN time preference",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 361,
+ 682,
+ 380,
+ 694
+ ],
+ "score": 0.92,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 381,
+ 681,
+ 505,
+ 694
+ ],
+ "score": 1.0,
+ "content": "is bell-shaped and varies from",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 106,
+ 692,
+ 339,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 692,
+ 339,
+ 705
+ ],
+ "score": 1.0,
+ "content": "2/3 at the ends to 1 at the middle frame (see Appendix E).",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 49
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 709,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Figs 6 and 9 show examples from the three settings. TAP approaches successfully discover interesting",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 719,
+ 505,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 719,
+ 505,
+ 734
+ ],
+ "score": 1.0,
+ "content": "bottlenecks in each setting. For grasping (Fig 6 (left)), both MIN and GENMIN consistently produce",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ }
+ ],
+ "index": 52.5
+ }
+ ],
+ "page_idx": 5,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 310,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 310,
+ 762
+ ],
+ "score": 1.0,
+ "content": "6",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 106,
+ 72,
+ 378,
+ 182
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 106,
+ 72,
+ 378,
+ 182
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 378,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 378,
+ 182
+ ],
+ "score": 0.968,
+ "type": "image",
+ "image_path": "5770695687526489da2068203ed0e833315b117d7b1003431ea451379f1cd9f3.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 378,
+ 108.66666666666666
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 108.66666666666666,
+ 378,
+ 145.33333333333331
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 145.33333333333331,
+ 378,
+ 181.99999999999997
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 191,
+ 378,
+ 211
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 190,
+ 379,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 190,
+ 379,
+ 202
+ ],
+ "score": 1.0,
+ "content": "Figure 3: (Best seen in pdf) One sample episode each for grasping, pick-and-",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 201,
+ 286,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 201,
+ 286,
+ 212
+ ],
+ "score": 1.0,
+ "content": "place, and pushing. Time overlaid on each frame.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5
+ }
+ ],
+ "index": 3.25
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 390,
+ 85,
+ 498,
+ 159
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 390,
+ 85,
+ 498,
+ 159
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 390,
+ 85,
+ 498,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 390,
+ 85,
+ 498,
+ 159
+ ],
+ "score": 0.943,
+ "type": "image",
+ "image_path": "c276a2df0e0bfd87996b9ced2d4c7e08dda70fc8eb2ef52b070ed396f8e3d0f5.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 390,
+ 85,
+ 498,
+ 122.0
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 390,
+ 122.0,
+ 498,
+ 159.0
+ ],
+ "spans": [],
+ "index": 3
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 387,
+ 160,
+ 505,
+ 210
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 387,
+ 160,
+ 504,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 387,
+ 160,
+ 495,
+ 171
+ ],
+ "score": 1.0,
+ "content": "Figure 4: Forward prediction",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 495,
+ 160,
+ 504,
+ 170
+ ],
+ "score": 0.81,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 387,
+ 170,
+ 505,
+ 181
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 387,
+ 170,
+ 505,
+ 181
+ ],
+ "score": 1.0,
+ "content": "error for grasping. TAP methods",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 387,
+ 180,
+ 506,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 387,
+ 180,
+ 506,
+ 190
+ ],
+ "score": 1.0,
+ "content": "(red) perform better than fixed-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 387,
+ 190,
+ 505,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 387,
+ 190,
+ 505,
+ 201
+ ],
+ "score": 1.0,
+ "content": "time predictors over all time",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 386,
+ 200,
+ 411,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 386,
+ 200,
+ 411,
+ 213
+ ],
+ "score": 1.0,
+ "content": "steps.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 5.75
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 224,
+ 200,
+ 236
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 223,
+ 201,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 223,
+ 201,
+ 238
+ ],
+ "score": 1.0,
+ "content": "4 EXPERIMENTS",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 247,
+ 505,
+ 314
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 248,
+ 505,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 248,
+ 505,
+ 259
+ ],
+ "score": 1.0,
+ "content": "We have proposed a time-agnostic prediction (TAP) paradigm that is different from the fixed-time",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 258,
+ 506,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 506,
+ 272
+ ],
+ "score": 1.0,
+ "content": "paradigm followed in prior prediction work. In our experiments, we focus on comparing TAP against",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 104,
+ 269,
+ 505,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 269,
+ 505,
+ 282
+ ],
+ "score": 1.0,
+ "content": "a representative fixed-time prediction model, keeping network architectures fixed. We use three",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 280,
+ 507,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 280,
+ 325,
+ 294
+ ],
+ "score": 1.0,
+ "content": "simulated robot manipulation settings: object grasping (",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 325,
+ 281,
+ 341,
+ 291
+ ],
+ "score": 0.41,
+ "content": "5 0 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 342,
+ 280,
+ 447,
+ 294
+ ],
+ "score": 1.0,
+ "content": "episodes), pick-and-place",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 447,
+ 281,
+ 464,
+ 291
+ ],
+ "score": 0.44,
+ "content": "7 5 \\mathrm { k }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 464,
+ 280,
+ 507,
+ 294
+ ],
+ "score": 1.0,
+ "content": "episodes),",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 303
+ ],
+ "score": 1.0,
+ "content": "and multi-object pushing (55k episodes). Example episodes from each task are shown in Fig 3 (videos",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 302,
+ 398,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 302,
+ 145,
+ 316
+ ],
+ "score": 1.0,
+ "content": "in Supp).",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 303,
+ 161,
+ 313
+ ],
+ "score": 0.79,
+ "content": "5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 161,
+ 302,
+ 329,
+ 316
+ ],
+ "score": 1.0,
+ "content": "of the data is set aside for testing. We use",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 330,
+ 303,
+ 364,
+ 313
+ ],
+ "score": 0.89,
+ "content": "6 4 \\times 6 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 364,
+ 302,
+ 398,
+ 316
+ ],
+ "score": 1.0,
+ "content": "images.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 15.5,
+ "bbox_fs": [
+ 104,
+ 248,
+ 507,
+ 316
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 319,
+ 505,
+ 374
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 320,
+ 505,
+ 332
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 320,
+ 505,
+ 332
+ ],
+ "score": 1.0,
+ "content": "For grasping (15 frames per episode), the arm moves to a single object on a table, selects a grasp, and",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 329,
+ 506,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 329,
+ 506,
+ 343
+ ],
+ "score": 1.0,
+ "content": "lifts it vertically. For pick-and-place (20 frames), the arm additionally places the lifted object at a new",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 340,
+ 506,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 340,
+ 506,
+ 354
+ ],
+ "score": 1.0,
+ "content": "position before performing a random walk. For pushing (40 frames), two objects are initialized at",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 351,
+ 505,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 505,
+ 366
+ ],
+ "score": 1.0,
+ "content": "random locations and pushed to random final positions. Object shapes and colors in all three settings",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 363,
+ 332,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 363,
+ 332,
+ 376
+ ],
+ "score": 1.0,
+ "content": "are randomly generated. Fig 3 shows example episodes.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 21,
+ "bbox_fs": [
+ 105,
+ 320,
+ 506,
+ 376
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 379,
+ 505,
+ 468
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 379,
+ 506,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 379,
+ 506,
+ 392
+ ],
+ "score": 1.0,
+ "content": "Forward prediction. First, we evaluate our approach for forward prediction on grasping. The first",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 389,
+ 506,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 389,
+ 506,
+ 404
+ ],
+ "score": 1.0,
+ "content": "frame (“start”) is provided as input. We train fixed-time baselines (architecture same as ours, using",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 402,
+ 506,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 402,
+ 116,
+ 412
+ ],
+ "score": 0.86,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 117,
+ 402,
+ 506,
+ 414
+ ],
+ "score": 1.0,
+ "content": "and GAN losses same as MIN and GENMIN) that target predictions at exactly 0.25, 0.50, 0.75, 1.0",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 412,
+ 505,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 412,
+ 505,
+ 425
+ ],
+ "score": 1.0,
+ "content": "fraction of the episode length (FIX0.25,. . . , FIX1.00). MIN and GENMIN are TAP with/without the",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 423,
+ 505,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 423,
+ 207,
+ 436
+ ],
+ "score": 1.0,
+ "content": "generalized minimum of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 208,
+ 424,
+ 238,
+ 434
+ ],
+ "score": 0.38,
+ "content": "{ \\tt S e c 3 . 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 239,
+ 423,
+ 505,
+ 436
+ ],
+ "score": 1.0,
+ "content": ". For GENMIN, we evaluate different choices of the time preference",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 433,
+ 504,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 433,
+ 134,
+ 447
+ ],
+ "score": 1.0,
+ "content": "vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 135,
+ 434,
+ 154,
+ 446
+ ],
+ "score": 0.91,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 155,
+ 433,
+ 230,
+ 447
+ ],
+ "score": 1.0,
+ "content": "(Sec 3.2). We set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 231,
+ 434,
+ 303,
+ 446
+ ],
+ "score": 0.92,
+ "content": "w ( t ) = \\beta + t / 1 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 303,
+ 433,
+ 496,
+ 447
+ ],
+ "score": 1.0,
+ "content": ", so that our preference increases linearly from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 496,
+ 435,
+ 504,
+ 446
+ ],
+ "score": 0.82,
+ "content": "\\beta",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 104,
+ 444,
+ 506,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 444,
+ 117,
+ 458
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 117,
+ 446,
+ 142,
+ 456
+ ],
+ "score": 0.89,
+ "content": "\\beta + 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 142,
+ 444,
+ 173,
+ 458
+ ],
+ "score": 1.0,
+ "content": ". Since",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 173,
+ 445,
+ 193,
+ 457
+ ],
+ "score": 0.91,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 194,
+ 444,
+ 313,
+ 458
+ ],
+ "score": 1.0,
+ "content": "applies multiplicatively, low",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 313,
+ 446,
+ 321,
+ 457
+ ],
+ "score": 0.83,
+ "content": "\\beta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 322,
+ 444,
+ 506,
+ 458
+ ],
+ "score": 1.0,
+ "content": "corresponds to high disparity in preferences",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 456,
+ 498,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 457,
+ 109,
+ 469
+ ],
+ "score": 0.0,
+ "content": "",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 110,
+ 457,
+ 140,
+ 468
+ ],
+ "score": 0.89,
+ "content": "\\beta = \\infty",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 140,
+ 457,
+ 428,
+ 469
+ ],
+ "score": 1.0,
+ "content": "reduces to MIN, i.e., no time preference). GENMIN2 is our approach with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 429,
+ 456,
+ 454,
+ 468
+ ],
+ "score": 0.91,
+ "content": "\\beta = 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 455,
+ 457,
+ 498,
+ 469
+ ],
+ "score": 1.0,
+ "content": "and so on.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 27.5,
+ "bbox_fs": [
+ 104,
+ 379,
+ 506,
+ 469
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 473,
+ 505,
+ 539
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 473,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 473,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "Fig 5 shows example predictions from all methods for the grasping task. In terms of visual quality",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 484,
+ 506,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 484,
+ 506,
+ 497
+ ],
+ "score": 1.0,
+ "content": "of predictions and finding a semantically coherent bottleneck, GENMIN2, GENMIN4, and GENMIN7",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 104,
+ 495,
+ 506,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 495,
+ 506,
+ 508
+ ],
+ "score": 1.0,
+ "content": "perform best—they reliably produce a grasp on the object while it is still on the table. With little or",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 506,
+ 505,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 506,
+ 505,
+ 519
+ ],
+ "score": 1.0,
+ "content": "no time preferences, MIN and GENMIN10 produce images very close to the start, while GENMIN0.5",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 517,
+ 506,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 517,
+ 506,
+ 530
+ ],
+ "score": 1.0,
+ "content": "places too high a value on predictions farther away, and produces blurry images of the object after",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 525,
+ 137,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 525,
+ 137,
+ 542
+ ],
+ "score": 1.0,
+ "content": "lifting.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 34.5,
+ "bbox_fs": [
+ 104,
+ 473,
+ 506,
+ 542
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 544,
+ 505,
+ 644
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 545,
+ 505,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 545,
+ 408,
+ 556
+ ],
+ "score": 1.0,
+ "content": "Quantitatively, for each method, we report the min and arg min index of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 408,
+ 545,
+ 418,
+ 556
+ ],
+ "score": 0.87,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 418,
+ 545,
+ 505,
+ 556
+ ],
+ "score": 1.0,
+ "content": "distance to all frames",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 505,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 192,
+ 569
+ ],
+ "score": 1.0,
+ "content": "in the video, as “min",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 193,
+ 556,
+ 203,
+ 567
+ ],
+ "score": 0.85,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 555,
+ 505,
+ 569
+ ],
+ "score": 1.0,
+ "content": "err” and “match-step” (“which future ground truth frame is the prediction",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 567,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 567,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": "closest to?”). Fig 4 shows a scatter plot, where each dot or square is one model. TAP (various models",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 578,
+ 506,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 578,
+ 160,
+ 591
+ ],
+ "score": 1.0,
+ "content": "with varying",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 161,
+ 578,
+ 168,
+ 590
+ ],
+ "score": 0.75,
+ "content": "\\beta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 169,
+ 578,
+ 506,
+ 591
+ ],
+ "score": 1.0,
+ "content": ") produces an even larger variation in stepsizes than fixed-time methods explicitly",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 589,
+ 506,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 589,
+ 215,
+ 601
+ ],
+ "score": 1.0,
+ "content": "targeting the entire video",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 216,
+ 589,
+ 248,
+ 599
+ ],
+ "score": 0.48,
+ "content": "\\mathrm { ' } _ { \\mathrm { F I X 0 . 7 5 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 249,
+ 589,
+ 267,
+ 601
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 268,
+ 589,
+ 295,
+ 599
+ ],
+ "score": 0.25,
+ "content": "\\mathrm { F I X } 1 . 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 295,
+ 589,
+ 506,
+ 601
+ ],
+ "score": 1.0,
+ "content": "fall short of producing predictions at 0.75 and 1.0",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 599,
+ 506,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 506,
+ 613
+ ],
+ "score": 1.0,
+ "content": "fraction of the episode length). TAP also produces higher quality predictions (lower error) over that",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 610,
+ 505,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 610,
+ 505,
+ 624
+ ],
+ "score": 1.0,
+ "content": "entire range. From these quantitative and qualitative results, we see that TAP not only successfully",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 621,
+ 505,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 621,
+ 505,
+ 635
+ ],
+ "score": 1.0,
+ "content": "encourages semantically coherent bottleneck predictions, it also produces higher quality predictions",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 633,
+ 326,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 326,
+ 645
+ ],
+ "score": 1.0,
+ "content": "than fixed-time prediction over a range of time offsets.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 42,
+ "bbox_fs": [
+ 105,
+ 545,
+ 506,
+ 645
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 649,
+ 505,
+ 704
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 505,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 505,
+ 662
+ ],
+ "score": 1.0,
+ "content": "Intermediate frame prediction. Next, we evaluate our approaches for bidirectionally conditioned",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 660,
+ 505,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 505,
+ 672
+ ],
+ "score": 1.0,
+ "content": "prediction in all three settings. Initial and final frames are provided as input, and the method is trained",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 505,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 505,
+ 684
+ ],
+ "score": 1.0,
+ "content": "to generate an intermediate frame. The FIX baseline now targets the middle frame. As before, MIN",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 681,
+ 505,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 681,
+ 360,
+ 694
+ ],
+ "score": 1.0,
+ "content": "and GENMIN are our TAP models. The GENMIN time preference",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 361,
+ 682,
+ 380,
+ 694
+ ],
+ "score": 0.92,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 381,
+ 681,
+ 505,
+ 694
+ ],
+ "score": 1.0,
+ "content": "is bell-shaped and varies from",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 106,
+ 692,
+ 339,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 692,
+ 339,
+ 705
+ ],
+ "score": 1.0,
+ "content": "2/3 at the ends to 1 at the middle frame (see Appendix E).",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 49,
+ "bbox_fs": [
+ 105,
+ 648,
+ 505,
+ 705
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 709,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Figs 6 and 9 show examples from the three settings. TAP approaches successfully discover interesting",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 719,
+ 505,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 719,
+ 505,
+ 734
+ ],
+ "score": 1.0,
+ "content": "bottlenecks in each setting. For grasping (Fig 6 (left)), both MIN and GENMIN consistently produce",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 105,
+ 399,
+ 506,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 399,
+ 506,
+ 414
+ ],
+ "score": 1.0,
+ "content": "clear images of the arm at the point at which it picks up the object. Pick-and-place (Fig 6, right) is",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 104,
+ 410,
+ 506,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 410,
+ 506,
+ 425
+ ],
+ "score": 1.0,
+ "content": "harder because it is more temporally extended, and the goal image does not specify how to grasp",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 422,
+ 505,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 422,
+ 505,
+ 435
+ ],
+ "score": 1.0,
+ "content": "the object. FIX struggles to produce any coherent predictions, but GENMIN once again identifies",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 432,
+ 506,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 432,
+ 506,
+ 447
+ ],
+ "score": 1.0,
+ "content": "bottlenecks reliably—in examples #3 and #1, it predicts the “pick” and the “place” respectively. For",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 444,
+ 505,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 444,
+ 505,
+ 456
+ ],
+ "score": 1.0,
+ "content": "the pushing setting (Fig 9 (left)), GENMIN frequently produces images with one object moved and the",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 455,
+ 506,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 455,
+ 506,
+ 468
+ ],
+ "score": 1.0,
+ "content": "other object fixed in place, which again is a semantically coherent bottleneck for this task. In row #1,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 466,
+ 448,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 448,
+ 478
+ ],
+ "score": 1.0,
+ "content": "it moves the correct object first to generate the subgoal, so that objects do not collide.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 52.5,
+ "bbox_fs": [
+ 105,
+ 708,
+ 505,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 115,
+ 79,
+ 495,
+ 202
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 115,
+ 79,
+ 495,
+ 202
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 79,
+ 495,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 79,
+ 495,
+ 202
+ ],
+ "score": 0.933,
+ "type": "image",
+ "image_path": "f86ef90a00ab7b85808ffe17668e41b8c711bf42aeb798ea3b554ada01459605.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 115,
+ 79,
+ 495,
+ 120.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 115,
+ 120.0,
+ 495,
+ 161.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 115,
+ 161.0,
+ 495,
+ 202.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 108,
+ 243,
+ 502,
+ 345
+ ],
+ "blocks": [
+ {
+ "type": "image_caption",
+ "bbox": [
+ 107,
+ 212,
+ 503,
+ 242
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 505,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 505,
+ 224
+ ],
+ "score": 1.0,
+ "content": "Figure 5: (Best seen in pdf) Forward prediction results on grasping comparing fixed-time predictors and our",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 222,
+ 505,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 222,
+ 505,
+ 234
+ ],
+ "score": 1.0,
+ "content": "approach. Each row is a separate example. First column is the input. Thereafter, each column corresponds to the",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 232,
+ 376,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 232,
+ 376,
+ 244
+ ],
+ "score": 1.0,
+ "content": "output of a different model per the column title. More in Appendix Fig 17.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "image_body",
+ "bbox": [
+ 108,
+ 243,
+ 502,
+ 345
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 243,
+ 502,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 243,
+ 502,
+ 345
+ ],
+ "score": 0.954,
+ "type": "image",
+ "image_path": "e5accb667c317aea1ebb4ab176afeda30572430c2be8e635f7851878016d9e41.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 108,
+ 243,
+ 502,
+ 277.0
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 108,
+ 277.0,
+ 502,
+ 311.0
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 108,
+ 311.0,
+ 502,
+ 345.0
+ ],
+ "spans": [],
+ "index": 8
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 355,
+ 505,
+ 396
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 355,
+ 506,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 355,
+ 506,
+ 367
+ ],
+ "score": 1.0,
+ "content": "Figure 6: (Best seen in pdf) Bidirectional prediction results comparing fixed-time prediction and our approach.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 366,
+ 505,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 366,
+ 505,
+ 376
+ ],
+ "score": 1.0,
+ "content": "(Left) Grasping results. First two columns are inputs (start and goal). Thereafter, each column corresponds to",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 374,
+ 505,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 374,
+ 505,
+ 387
+ ],
+ "score": 1.0,
+ "content": "the output of a different model per the column title. “match” is the ground truth image closest to the GENMIN",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 384,
+ 496,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 496,
+ 398
+ ],
+ "score": 1.0,
+ "content": "prediction. More in Appendix Fig 13. (Right) Similar results for pick-and-place. More in Appendix Fig 14.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 10.5
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 400,
+ 505,
+ 478
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 399,
+ 506,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 399,
+ 506,
+ 414
+ ],
+ "score": 1.0,
+ "content": "clear images of the arm at the point at which it picks up the object. Pick-and-place (Fig 6, right) is",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 104,
+ 410,
+ 506,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 410,
+ 506,
+ 425
+ ],
+ "score": 1.0,
+ "content": "harder because it is more temporally extended, and the goal image does not specify how to grasp",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 422,
+ 505,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 422,
+ 505,
+ 435
+ ],
+ "score": 1.0,
+ "content": "the object. FIX struggles to produce any coherent predictions, but GENMIN once again identifies",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 432,
+ 506,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 432,
+ 506,
+ 447
+ ],
+ "score": 1.0,
+ "content": "bottlenecks reliably—in examples #3 and #1, it predicts the “pick” and the “place” respectively. For",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 444,
+ 505,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 444,
+ 505,
+ 456
+ ],
+ "score": 1.0,
+ "content": "the pushing setting (Fig 9 (left)), GENMIN frequently produces images with one object moved and the",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 455,
+ 506,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 455,
+ 506,
+ 468
+ ],
+ "score": 1.0,
+ "content": "other object fixed in place, which again is a semantically coherent bottleneck for this task. In row #1,",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 466,
+ 448,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 448,
+ 478
+ ],
+ "score": 1.0,
+ "content": "it moves the correct object first to generate the subgoal, so that objects do not collide.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 604
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 494,
+ 495
+ ],
+ "score": 1.0,
+ "content": "Table 1 shows quantitative results over the full test set. As in forward prediction, we report min",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 495,
+ 483,
+ 505,
+ 494
+ ],
+ "score": 0.85,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 494,
+ 505,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 505,
+ 506
+ ],
+ "score": 1.0,
+ "content": "error and the best-matching frame index (“match-step”) for all methods. MIN and GENMIN consistently",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 104,
+ 505,
+ 506,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 505,
+ 506,
+ 518
+ ],
+ "score": 1.0,
+ "content": "yield higher quality predictions (lower error) than FIX at similar times on average. As an example,",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 516,
+ 506,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 516,
+ 227,
+ 528
+ ],
+ "score": 1.0,
+ "content": "GENMIN reduces FIX errors by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 227,
+ 516,
+ 247,
+ 527
+ ],
+ "score": 0.84,
+ "content": "21 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 247,
+ 516,
+ 250,
+ 528
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 251,
+ 516,
+ 278,
+ 527
+ ],
+ "score": 0.86,
+ "content": "2 6 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 278,
+ 516,
+ 298,
+ 528
+ ],
+ "score": 1.0,
+ "content": ", and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 299,
+ 516,
+ 326,
+ 527
+ ],
+ "score": 0.87,
+ "content": "5 3 . 2 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 327,
+ 516,
+ 506,
+ 528
+ ],
+ "score": 1.0,
+ "content": "on the three tasks—these are consistent and",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 104,
+ 526,
+ 506,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 526,
+ 506,
+ 541
+ ],
+ "score": 1.0,
+ "content": "large gains that increase with increasing task complexity/duration. Additionally, while all foregoing",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 506,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 506,
+ 550
+ ],
+ "score": 1.0,
+ "content": "results were reported without the CVAE approach of Sec 3.4, Table 1 shows results for GENMIN+VAE,",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 549,
+ 505,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 549,
+ 505,
+ 561
+ ],
+ "score": 1.0,
+ "content": "and Fig 9 shows example predictions for pick-and-place. In our models, individual stochastic",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 560,
+ 505,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 209,
+ 572
+ ],
+ "score": 1.0,
+ "content": "predictions from GENMIN",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 209,
+ 562,
+ 216,
+ 570
+ ],
+ "score": 0.47,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 217,
+ 560,
+ 297,
+ 572
+ ],
+ "score": 1.0,
+ "content": "VAE produce higher",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 297,
+ 560,
+ 307,
+ 571
+ ],
+ "score": 0.87,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 308,
+ 560,
+ 505,
+ 572
+ ],
+ "score": 1.0,
+ "content": "errors than GENMIN. However, the CVAE helps",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 570,
+ 506,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 570,
+ 506,
+ 583
+ ],
+ "score": 1.0,
+ "content": "capture meaningful sources of stochasticity at the bottlenecks—in Fig 9 (right), it produces different",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 594
+ ],
+ "score": 1.0,
+ "content": "grasp configurations to pick up the object in each case. To measure this, we evaluate the best of 100",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 592,
+ 505,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 592,
+ 248,
+ 605
+ ],
+ "score": 1.0,
+ "content": "stochastic predictions from GENMIN",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 248,
+ 595,
+ 254,
+ 603
+ ],
+ "score": 0.42,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 255,
+ 592,
+ 347,
+ 605
+ ],
+ "score": 1.0,
+ "content": "VAE in Table 1 (GENMIN",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 348,
+ 595,
+ 354,
+ 602
+ ],
+ "score": 0.31,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 354,
+ 592,
+ 505,
+ 605
+ ],
+ "score": 1.0,
+ "content": "VAE BEST-OF-100). On pick-and-place",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 108,
+ 623,
+ 496,
+ 707
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 108,
+ 623,
+ 496,
+ 707
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 623,
+ 496,
+ 707
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 623,
+ 496,
+ 707
+ ],
+ "score": 0.976,
+ "html": "
| Setting→ Method↓ | Grasping (T=15 steps) min lierr | match-step/T | Pick-and-place (T=20 steps) min l1 err | match-step/T | Pushing (T=30 steps) min l1 err | match-step/T |
| fix | 0.0153 | 0.51±0.17 | 0.0366 | 0.53±0.24 | 0.0722 | 0.36±0.19 |
| MIN (ours) | 0.0104 | 0.49±0.18 | 0.0256 | 0.41±0.30 | 0.0365 | 0.35±0.22 |
| GENMIN (ours) | 0.0121 | 0.45±0.16 | 0.0269 | 0.42±0.23 | 0.0338 | 0.36±0.19 |
| GENMIN W/O GAN (Ours) | 0.0117 | 0.45±0.16 | 0.0235 | 0.46±0.25 | 0.0411 | 0.37±0.19 |
| GENMIN + VAE (Ours) | 0.0156 | 0.47±0.18 | 0.0432 | 0.31±0.24 | 0.0447 | 0.37±0.21 |
| GENMIN + VAE BEST-OF-10O (Ours) | 0.0121 | - | 0.0196 | - | 0.0236 | - |
",
+ "type": "table",
+ "image_path": "1084ec4c5632b944c87b5533346157f4ba859696b63576a0934c34e863408242.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 32,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 108,
+ 623,
+ 496,
+ 651.0
+ ],
+ "spans": [],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 108,
+ 651.0,
+ 496,
+ 679.0
+ ],
+ "spans": [],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 108,
+ 679.0,
+ 496,
+ 707.0
+ ],
+ "spans": [],
+ "index": 33
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 107,
+ 715,
+ 504,
+ 746
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 714,
+ 505,
+ 728
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 714,
+ 505,
+ 728
+ ],
+ "score": 1.0,
+ "content": "Table 1: Bidirectional frame prediction performance on: grasping, pick-and-place, and two-object pushing.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 725,
+ 505,
+ 737
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 725,
+ 148,
+ 737
+ ],
+ "score": 1.0,
+ "content": "Lower min",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 149,
+ 726,
+ 158,
+ 735
+ ],
+ "score": 0.85,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 158,
+ 725,
+ 421,
+ 737
+ ],
+ "score": 1.0,
+ "content": "err is better. match-step denotes which times are being predicted, out of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 421,
+ 726,
+ 429,
+ 735
+ ],
+ "score": 0.74,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 429,
+ 725,
+ 505,
+ 737
+ ],
+ "score": 1.0,
+ "content": "steps. It is clear that",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 736,
+ 428,
+ 746
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 736,
+ 428,
+ 746
+ ],
+ "score": 1.0,
+ "content": "TAP methods make better predictions than fixed-time prediction at the same time offsets.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 33.5
+ }
+ ],
+ "page_idx": 6,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "7",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 115,
+ 79,
+ 495,
+ 202
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 115,
+ 79,
+ 495,
+ 202
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 79,
+ 495,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 79,
+ 495,
+ 202
+ ],
+ "score": 0.933,
+ "type": "image",
+ "image_path": "f86ef90a00ab7b85808ffe17668e41b8c711bf42aeb798ea3b554ada01459605.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 115,
+ 79,
+ 495,
+ 120.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 115,
+ 120.0,
+ 495,
+ 161.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 115,
+ 161.0,
+ 495,
+ 202.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 108,
+ 243,
+ 502,
+ 345
+ ],
+ "blocks": [
+ {
+ "type": "image_caption",
+ "bbox": [
+ 107,
+ 212,
+ 503,
+ 242
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 505,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 505,
+ 224
+ ],
+ "score": 1.0,
+ "content": "Figure 5: (Best seen in pdf) Forward prediction results on grasping comparing fixed-time predictors and our",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 222,
+ 505,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 222,
+ 505,
+ 234
+ ],
+ "score": 1.0,
+ "content": "approach. Each row is a separate example. First column is the input. Thereafter, each column corresponds to the",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 232,
+ 376,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 232,
+ 376,
+ 244
+ ],
+ "score": 1.0,
+ "content": "output of a different model per the column title. More in Appendix Fig 17.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "image_body",
+ "bbox": [
+ 108,
+ 243,
+ 502,
+ 345
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 243,
+ 502,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 243,
+ 502,
+ 345
+ ],
+ "score": 0.954,
+ "type": "image",
+ "image_path": "e5accb667c317aea1ebb4ab176afeda30572430c2be8e635f7851878016d9e41.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 108,
+ 243,
+ 502,
+ 277.0
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 108,
+ 277.0,
+ 502,
+ 311.0
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 108,
+ 311.0,
+ 502,
+ 345.0
+ ],
+ "spans": [],
+ "index": 8
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 355,
+ 505,
+ 396
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 355,
+ 506,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 355,
+ 506,
+ 367
+ ],
+ "score": 1.0,
+ "content": "Figure 6: (Best seen in pdf) Bidirectional prediction results comparing fixed-time prediction and our approach.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 366,
+ 505,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 366,
+ 505,
+ 376
+ ],
+ "score": 1.0,
+ "content": "(Left) Grasping results. First two columns are inputs (start and goal). Thereafter, each column corresponds to",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 374,
+ 505,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 374,
+ 505,
+ 387
+ ],
+ "score": 1.0,
+ "content": "the output of a different model per the column title. “match” is the ground truth image closest to the GENMIN",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 384,
+ 496,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 384,
+ 496,
+ 398
+ ],
+ "score": 1.0,
+ "content": "prediction. More in Appendix Fig 13. (Right) Similar results for pick-and-place. More in Appendix Fig 14.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 10.5
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 400,
+ 505,
+ 478
+ ],
+ "lines": [],
+ "index": 16,
+ "bbox_fs": [
+ 104,
+ 399,
+ 506,
+ 478
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 604
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 494,
+ 495
+ ],
+ "score": 1.0,
+ "content": "Table 1 shows quantitative results over the full test set. As in forward prediction, we report min",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 495,
+ 483,
+ 505,
+ 494
+ ],
+ "score": 0.85,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 494,
+ 505,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 505,
+ 506
+ ],
+ "score": 1.0,
+ "content": "error and the best-matching frame index (“match-step”) for all methods. MIN and GENMIN consistently",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 104,
+ 505,
+ 506,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 505,
+ 506,
+ 518
+ ],
+ "score": 1.0,
+ "content": "yield higher quality predictions (lower error) than FIX at similar times on average. As an example,",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 516,
+ 506,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 516,
+ 227,
+ 528
+ ],
+ "score": 1.0,
+ "content": "GENMIN reduces FIX errors by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 227,
+ 516,
+ 247,
+ 527
+ ],
+ "score": 0.84,
+ "content": "21 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 247,
+ 516,
+ 250,
+ 528
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 251,
+ 516,
+ 278,
+ 527
+ ],
+ "score": 0.86,
+ "content": "2 6 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 278,
+ 516,
+ 298,
+ 528
+ ],
+ "score": 1.0,
+ "content": ", and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 299,
+ 516,
+ 326,
+ 527
+ ],
+ "score": 0.87,
+ "content": "5 3 . 2 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 327,
+ 516,
+ 506,
+ 528
+ ],
+ "score": 1.0,
+ "content": "on the three tasks—these are consistent and",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 104,
+ 526,
+ 506,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 526,
+ 506,
+ 541
+ ],
+ "score": 1.0,
+ "content": "large gains that increase with increasing task complexity/duration. Additionally, while all foregoing",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 506,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 506,
+ 550
+ ],
+ "score": 1.0,
+ "content": "results were reported without the CVAE approach of Sec 3.4, Table 1 shows results for GENMIN+VAE,",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 549,
+ 505,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 549,
+ 505,
+ 561
+ ],
+ "score": 1.0,
+ "content": "and Fig 9 shows example predictions for pick-and-place. In our models, individual stochastic",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 560,
+ 505,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 209,
+ 572
+ ],
+ "score": 1.0,
+ "content": "predictions from GENMIN",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 209,
+ 562,
+ 216,
+ 570
+ ],
+ "score": 0.47,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 217,
+ 560,
+ 297,
+ 572
+ ],
+ "score": 1.0,
+ "content": "VAE produce higher",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 297,
+ 560,
+ 307,
+ 571
+ ],
+ "score": 0.87,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 308,
+ 560,
+ 505,
+ 572
+ ],
+ "score": 1.0,
+ "content": "errors than GENMIN. However, the CVAE helps",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 570,
+ 506,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 570,
+ 506,
+ 583
+ ],
+ "score": 1.0,
+ "content": "capture meaningful sources of stochasticity at the bottlenecks—in Fig 9 (right), it produces different",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 594
+ ],
+ "score": 1.0,
+ "content": "grasp configurations to pick up the object in each case. To measure this, we evaluate the best of 100",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 592,
+ 505,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 592,
+ 248,
+ 605
+ ],
+ "score": 1.0,
+ "content": "stochastic predictions from GENMIN",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 248,
+ 595,
+ 254,
+ 603
+ ],
+ "score": 0.42,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 255,
+ 592,
+ 347,
+ 605
+ ],
+ "score": 1.0,
+ "content": "VAE in Table 1 (GENMIN",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 348,
+ 595,
+ 354,
+ 602
+ ],
+ "score": 0.31,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 354,
+ 592,
+ 505,
+ 605
+ ],
+ "score": 1.0,
+ "content": "VAE BEST-OF-100). On pick-and-place",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 253,
+ 506,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 253,
+ 506,
+ 265
+ ],
+ "score": 1.0,
+ "content": "and pushing, the best VAE results are significantly better than any of the deterministic methods.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 263,
+ 505,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 263,
+ 468,
+ 276
+ ],
+ "score": 1.0,
+ "content": "Table 1 also shows results for our method without the GAN (GENMIN W/O GAN)—while its",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 468,
+ 264,
+ 478,
+ 275
+ ],
+ "score": 0.88,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 478,
+ 263,
+ 505,
+ 276
+ ],
+ "score": 1.0,
+ "content": "errors",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 275,
+ 321,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 275,
+ 321,
+ 288
+ ],
+ "score": 1.0,
+ "content": "are comparable, we observed a drop in visual quality.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 25,
+ "bbox_fs": [
+ 104,
+ 482,
+ 506,
+ 605
+ ]
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 108,
+ 623,
+ 496,
+ 707
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 108,
+ 623,
+ 496,
+ 707
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 623,
+ 496,
+ 707
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 623,
+ 496,
+ 707
+ ],
+ "score": 0.976,
+ "html": "| Setting→ Method↓ | Grasping (T=15 steps) min lierr | match-step/T | Pick-and-place (T=20 steps) min l1 err | match-step/T | Pushing (T=30 steps) min l1 err | match-step/T |
| fix | 0.0153 | 0.51±0.17 | 0.0366 | 0.53±0.24 | 0.0722 | 0.36±0.19 |
| MIN (ours) | 0.0104 | 0.49±0.18 | 0.0256 | 0.41±0.30 | 0.0365 | 0.35±0.22 |
| GENMIN (ours) | 0.0121 | 0.45±0.16 | 0.0269 | 0.42±0.23 | 0.0338 | 0.36±0.19 |
| GENMIN W/O GAN (Ours) | 0.0117 | 0.45±0.16 | 0.0235 | 0.46±0.25 | 0.0411 | 0.37±0.19 |
| GENMIN + VAE (Ours) | 0.0156 | 0.47±0.18 | 0.0432 | 0.31±0.24 | 0.0447 | 0.37±0.21 |
| GENMIN + VAE BEST-OF-10O (Ours) | 0.0121 | - | 0.0196 | - | 0.0236 | - |
",
+ "type": "table",
+ "image_path": "1084ec4c5632b944c87b5533346157f4ba859696b63576a0934c34e863408242.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 32,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 108,
+ 623,
+ 496,
+ 651.0
+ ],
+ "spans": [],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 108,
+ 651.0,
+ 496,
+ 679.0
+ ],
+ "spans": [],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 108,
+ 679.0,
+ 496,
+ 707.0
+ ],
+ "spans": [],
+ "index": 33
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 107,
+ 715,
+ 504,
+ 746
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 714,
+ 505,
+ 728
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 714,
+ 505,
+ 728
+ ],
+ "score": 1.0,
+ "content": "Table 1: Bidirectional frame prediction performance on: grasping, pick-and-place, and two-object pushing.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 725,
+ 505,
+ 737
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 725,
+ 148,
+ 737
+ ],
+ "score": 1.0,
+ "content": "Lower min",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 149,
+ 726,
+ 158,
+ 735
+ ],
+ "score": 0.85,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 158,
+ 725,
+ 421,
+ 737
+ ],
+ "score": 1.0,
+ "content": "err is better. match-step denotes which times are being predicted, out of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 421,
+ 726,
+ 429,
+ 735
+ ],
+ "score": 0.74,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 429,
+ 725,
+ 505,
+ 737
+ ],
+ "score": 1.0,
+ "content": "steps. It is clear that",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 736,
+ 428,
+ 746
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 736,
+ 428,
+ 746
+ ],
+ "score": 1.0,
+ "content": "TAP methods make better predictions than fixed-time prediction at the same time offsets.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 33.5
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 108,
+ 79,
+ 502,
+ 187
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 108,
+ 79,
+ 502,
+ 187
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 79,
+ 502,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 79,
+ 502,
+ 187
+ ],
+ "score": 0.96,
+ "type": "image",
+ "image_path": "08ead4e71ea153e3db3ca86982af8338fef990eac09998525a0fdc8907203658.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 108,
+ 79,
+ 502,
+ 115.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 108,
+ 115.0,
+ 502,
+ 151.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 108,
+ 151.0,
+ 502,
+ 187.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 197,
+ 506,
+ 237
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 197,
+ 505,
+ 209
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 197,
+ 505,
+ 209
+ ],
+ "score": 1.0,
+ "content": "Figure 9: (Best seen in pdf) (Left) Bidirectional prediction results on two-object pushing. More in Appendix",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 207,
+ 506,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 207,
+ 506,
+ 218
+ ],
+ "score": 1.0,
+ "content": "Fig 15. (Right) When used with a VAE (Sec 3.4), our approach captures residual stochasticity at the bottleneck.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 216,
+ 506,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 216,
+ 302,
+ 228
+ ],
+ "score": 1.0,
+ "content": "In these results from the pick-and-place task, GENMIN",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 303,
+ 219,
+ 309,
+ 226
+ ],
+ "score": 0.37,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 309,
+ 216,
+ 506,
+ 228
+ ],
+ "score": 1.0,
+ "content": "VAE produces images that are all of the arm in contact",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 226,
+ 474,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 226,
+ 474,
+ 239
+ ],
+ "score": 1.0,
+ "content": "with the object on the table, but at different points on the object, and with different arm/gripper poses.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5
+ }
+ ],
+ "index": 2.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 253,
+ 506,
+ 286
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 253,
+ 506,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 253,
+ 506,
+ 265
+ ],
+ "score": 1.0,
+ "content": "and pushing, the best VAE results are significantly better than any of the deterministic methods.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 263,
+ 505,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 263,
+ 468,
+ 276
+ ],
+ "score": 1.0,
+ "content": "Table 1 also shows results for our method without the GAN (GENMIN W/O GAN)—while its",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 468,
+ 264,
+ 478,
+ 275
+ ],
+ "score": 0.88,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 478,
+ 263,
+ 505,
+ 276
+ ],
+ "score": 1.0,
+ "content": "errors",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 275,
+ 321,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 275,
+ 321,
+ 288
+ ],
+ "score": 1.0,
+ "content": "are comparable, we observed a drop in visual quality.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 292,
+ 304,
+ 412
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 304,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 167,
+ 304
+ ],
+ "score": 1.0,
+ "content": "As indicated in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 167,
+ 292,
+ 197,
+ 303
+ ],
+ "score": 0.25,
+ "content": "{ \\mathrm { S e c } } 3 . 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 197,
+ 291,
+ 304,
+ 304
+ ],
+ "score": 1.0,
+ "content": "and Eq 3, TAP may also be",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 303,
+ 305,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 303,
+ 305,
+ 315
+ ],
+ "score": 1.0,
+ "content": "applied recursively. Fig 7 compares consecutive",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 314,
+ 305,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 314,
+ 305,
+ 326
+ ],
+ "score": 1.0,
+ "content": "subgoals for the pick-and-place task produced by",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 325,
+ 306,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 325,
+ 306,
+ 335
+ ],
+ "score": 1.0,
+ "content": "recursive TAP versus a recursive fixed-time model.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 335,
+ 306,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 335,
+ 172,
+ 348
+ ],
+ "score": 1.0,
+ "content": "Recursion level",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 172,
+ 336,
+ 198,
+ 346
+ ],
+ "score": 0.92,
+ "content": "r = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 198,
+ 335,
+ 306,
+ 348
+ ],
+ "score": 1.0,
+ "content": "refers to the first subgoal,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 347,
+ 305,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 347,
+ 124,
+ 358
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 124,
+ 347,
+ 150,
+ 357
+ ],
+ "score": 0.89,
+ "content": "r = 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 150,
+ 347,
+ 305,
+ 358
+ ],
+ "score": 1.0,
+ "content": "refers to the subgoal generated when",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 357,
+ 305,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 357,
+ 305,
+ 370
+ ],
+ "score": 1.0,
+ "content": "the first subgoal is provided as the goal input (start",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 369,
+ 305,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 369,
+ 305,
+ 381
+ ],
+ "score": 1.0,
+ "content": "input is unchanged). In example #1, FIX struggles",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 379,
+ 306,
+ 391
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 379,
+ 306,
+ 391
+ ],
+ "score": 1.0,
+ "content": "while “ours” identifies the “place” bottleneck at",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 107,
+ 390,
+ 306,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 391,
+ 131,
+ 401
+ ],
+ "score": 0.88,
+ "content": "r = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 132,
+ 390,
+ 306,
+ 402
+ ],
+ "score": 1.0,
+ "content": ", and subsequently the “pick” botleneck at",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 402,
+ 135,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 402,
+ 131,
+ 412
+ ],
+ "score": 0.84,
+ "content": "r = 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 131,
+ 402,
+ 135,
+ 412
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 312,
+ 291,
+ 504,
+ 371
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 312,
+ 291,
+ 504,
+ 371
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 312,
+ 291,
+ 504,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 312,
+ 291,
+ 504,
+ 371
+ ],
+ "score": 0.969,
+ "type": "image",
+ "image_path": "fe2aa8133a51dbd41e95dc77f0fc36f12af4aaa53458e7768be0fa33c485273a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 23.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 312,
+ 291,
+ 504,
+ 304.3333333333333
+ ],
+ "spans": [],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 312,
+ 304.3333333333333,
+ 504,
+ 317.66666666666663
+ ],
+ "spans": [],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 312,
+ 317.66666666666663,
+ 504,
+ 330.99999999999994
+ ],
+ "spans": [],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 312,
+ 330.99999999999994,
+ 504,
+ 344.33333333333326
+ ],
+ "spans": [],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 312,
+ 344.33333333333326,
+ 504,
+ 357.6666666666666
+ ],
+ "spans": [],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 312,
+ 357.6666666666666,
+ 504,
+ 370.9999999999999
+ ],
+ "spans": [],
+ "index": 26
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 312,
+ 376,
+ 505,
+ 397
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 312,
+ 376,
+ 505,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 312,
+ 376,
+ 505,
+ 387
+ ],
+ "score": 1.0,
+ "content": "Figure 7: Recursive bidirectional prediction on pick-",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 312,
+ 387,
+ 481,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 312,
+ 387,
+ 352,
+ 397
+ ],
+ "score": 1.0,
+ "content": "and-place.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 352,
+ 387,
+ 375,
+ 396
+ ],
+ "score": 0.88,
+ "content": "r = 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 375,
+ 387,
+ 454,
+ 397
+ ],
+ "score": 1.0,
+ "content": "is earlier in time than",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 455,
+ 387,
+ 477,
+ 396
+ ],
+ "score": 0.88,
+ "content": "r = 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 477,
+ 387,
+ 481,
+ 397
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27.5
+ }
+ ],
+ "index": 25.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 418,
+ 354,
+ 550
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 417,
+ 356,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 417,
+ 356,
+ 431
+ ],
+ "score": 1.0,
+ "content": "Finally, we test on “BAIR pushing” (Ebert et al., 2017), a",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 429,
+ 355,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 355,
+ 442
+ ],
+ "score": 1.0,
+ "content": "real-world dataset that is commonly used in visual prediction",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 440,
+ 355,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 440,
+ 355,
+ 452
+ ],
+ "score": 1.0,
+ "content": "tasks. The data consists of 30-frame clips of random motions",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 451,
+ 355,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 451,
+ 355,
+ 463
+ ],
+ "score": 1.0,
+ "content": "of a Sawyer arm tabletop. While this dataset does not have",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 462,
+ 355,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 296,
+ 474
+ ],
+ "score": 1.0,
+ "content": "natural bottlenecks like in grasping, TAP (min",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 296,
+ 462,
+ 306,
+ 473
+ ],
+ "score": 0.84,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 307,
+ 462,
+ 355,
+ 474
+ ],
+ "score": 1.0,
+ "content": "error 0.046",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 473,
+ 356,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 473,
+ 356,
+ 485
+ ],
+ "score": 1.0,
+ "content": "at match-step 15.42) still performs better than FIX (0.059 at",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 484,
+ 356,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 484,
+ 356,
+ 496
+ ],
+ "score": 1.0,
+ "content": "15.29). Qualitatively, as Fig 8 shows, even though BAIR push-",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 495,
+ 356,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 356,
+ 507
+ ],
+ "score": 1.0,
+ "content": "ing contains incoherent random arm motions, TAP consistently",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 506,
+ 356,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 506,
+ 356,
+ 518
+ ],
+ "score": 1.0,
+ "content": "produces predictions that plausibly lie on the path from start to",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 517,
+ 355,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 517,
+ 355,
+ 529
+ ],
+ "score": 1.0,
+ "content": "goal image. In example #1, given a start and goal image with",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 528,
+ 355,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 528,
+ 355,
+ 540
+ ],
+ "score": 1.0,
+ "content": "one object displaced, “ours” correctly moves the arm to the",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 538,
+ 309,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 538,
+ 309,
+ 552
+ ],
+ "score": 1.0,
+ "content": "object before displacement, whereas FIX struggles.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 34.5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 362,
+ 418,
+ 504,
+ 495
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 362,
+ 418,
+ 504,
+ 495
+ ],
+ "group_id": 2,
+ "lines": [
+ {
+ "bbox": [
+ 362,
+ 418,
+ 504,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 362,
+ 418,
+ 504,
+ 495
+ ],
+ "score": 0.965,
+ "type": "image",
+ "image_path": "9385c328efd3ef9d40d2aebb6da428efee79f90d0d1504ea5d2217964752ce67.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 41.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 362,
+ 418,
+ 504,
+ 456.5
+ ],
+ "spans": [],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 362,
+ 456.5,
+ 504,
+ 495.0
+ ],
+ "spans": [],
+ "index": 42
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 362,
+ 502,
+ 505,
+ 542
+ ],
+ "group_id": 2,
+ "lines": [
+ {
+ "bbox": [
+ 361,
+ 501,
+ 506,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 361,
+ 501,
+ 506,
+ 513
+ ],
+ "score": 1.0,
+ "content": "Figure 8: Bidirectional prediction re-",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 362,
+ 512,
+ 505,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 362,
+ 512,
+ 505,
+ 523
+ ],
+ "score": 1.0,
+ "content": "sults on BAIR pushing data. The first",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 362,
+ 522,
+ 505,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 362,
+ 522,
+ 505,
+ 533
+ ],
+ "score": 1.0,
+ "content": "two columns are the inputs, and the next",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 361,
+ 532,
+ 495,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 361,
+ 532,
+ 495,
+ 542
+ ],
+ "score": 1.0,
+ "content": "two correspond to FIX and GENMIN.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 44.5
+ }
+ ],
+ "index": 43.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 567,
+ 396,
+ 687
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 567,
+ 396,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 567,
+ 396,
+ 579
+ ],
+ "score": 1.0,
+ "content": "Bottleneck discovery frequency. We have thus far relied on qualitative",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 578,
+ 396,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 578,
+ 396,
+ 590
+ ],
+ "score": 1.0,
+ "content": "results to assess how often our approach finds coherent bottlenecks. For",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 589,
+ 397,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 589,
+ 397,
+ 602
+ ],
+ "score": 1.0,
+ "content": "pushing, we test bottleneck discovery frequency more quantitatively. We",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 598,
+ 397,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 598,
+ 397,
+ 614
+ ],
+ "score": 1.0,
+ "content": "make the reasonable assumption that bottlenecks in 2-object pushing",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 106,
+ 611,
+ 398,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 611,
+ 398,
+ 623
+ ],
+ "score": 1.0,
+ "content": "correspond to states where one object is pushed and the other is in place.",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 621,
+ 397,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 397,
+ 634
+ ],
+ "score": 1.0,
+ "content": "Our metric exploits knowledge of true object positions at start and goal",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 106,
+ 633,
+ 396,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 396,
+ 644
+ ],
+ "score": 1.0,
+ "content": "states. First, for this evaluation, we restrict both GENMIN and FIX to",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 397,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 397,
+ 656
+ ],
+ "score": 1.0,
+ "content": "synthesize predictions purely by warping and masking inputs. Thus, we",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 105,
+ 655,
+ 397,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 655,
+ 397,
+ 667
+ ],
+ "score": 1.0,
+ "content": "can track where the pixels at ground truth object locations in start and",
+ "type": "text"
+ }
+ ],
+ "index": 57
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 397,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 397,
+ 678
+ ],
+ "score": 1.0,
+ "content": "goal images end up in the prediction, i.e., where did each object move?",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 396,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 396,
+ 689
+ ],
+ "score": 1.0,
+ "content": "We then compute a score that may be thresholded to detect when only",
+ "type": "text"
+ }
+ ],
+ "index": 60
+ }
+ ],
+ "index": 53
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 406,
+ 570,
+ 500,
+ 647
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 406,
+ 570,
+ 500,
+ 647
+ ],
+ "group_id": 3,
+ "lines": [
+ {
+ "bbox": [
+ 406,
+ 570,
+ 500,
+ 647
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 406,
+ 570,
+ 500,
+ 647
+ ],
+ "score": 0.958,
+ "type": "image",
+ "image_path": "c98be0a524e6a6ba7b94ce1b7a36d75a2305e1440f91fbf772d0f3efa5ba64f2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 53.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 406,
+ 570,
+ 500,
+ 608.5
+ ],
+ "spans": [],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 406,
+ 608.5,
+ 500,
+ 647.0
+ ],
+ "spans": [],
+ "index": 56
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 403,
+ 659,
+ 505,
+ 679
+ ],
+ "group_id": 3,
+ "lines": [
+ {
+ "bbox": [
+ 403,
+ 659,
+ 506,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 403,
+ 659,
+ 506,
+ 669
+ ],
+ "score": 1.0,
+ "content": "Figure 10: Bottleneck fre-",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 403,
+ 670,
+ 501,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 403,
+ 670,
+ 501,
+ 679
+ ],
+ "score": 1.0,
+ "content": "quency vs. score threshold",
+ "type": "text"
+ }
+ ],
+ "index": 61
+ }
+ ],
+ "index": 59.5
+ }
+ ],
+ "index": 56.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 689,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 701
+ ],
+ "score": 1.0,
+ "content": "one object moves (details in Appendix F). As an intuitive example, suppose that the two objects are",
+ "type": "text"
+ }
+ ],
+ "index": 62
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 159,
+ 711
+ ],
+ "score": 1.0,
+ "content": "displaced by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 160,
+ 699,
+ 186,
+ 709
+ ],
+ "score": 0.68,
+ "content": "1 0 \\mathrm { c m }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 186,
+ 699,
+ 203,
+ 711
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 204,
+ 699,
+ 230,
+ 709
+ ],
+ "score": 0.65,
+ "content": "1 5 \\mathrm { { c m } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 230,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "respectively between start and goal frames. Suppose further that our",
+ "type": "text"
+ }
+ ],
+ "index": 63
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "predictor predicts the start frame as its output. Then its distance score would be computed to be 10",
+ "type": "text"
+ }
+ ],
+ "index": 64
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 507,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 256,
+ 733
+ ],
+ "score": 1.0,
+ "content": "cm. For all values of threshold below",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 257,
+ 721,
+ 282,
+ 731
+ ],
+ "score": 0.67,
+ "content": "1 0 \\mathrm { c m }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 282,
+ 720,
+ 507,
+ 733
+ ],
+ "score": 1.0,
+ "content": ", this would be counted as a bottleneck discovery failure.",
+ "type": "text"
+ }
+ ],
+ "index": 65
+ }
+ ],
+ "index": 63.5
+ }
+ ],
+ "page_idx": 7,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 300,
+ 750,
+ 309,
+ 761
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 300,
+ 750,
+ 309,
+ 761
+ ],
+ "score": 1.0,
+ "content": "8",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 108,
+ 79,
+ 502,
+ 187
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 108,
+ 79,
+ 502,
+ 187
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 79,
+ 502,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 79,
+ 502,
+ 187
+ ],
+ "score": 0.96,
+ "type": "image",
+ "image_path": "08ead4e71ea153e3db3ca86982af8338fef990eac09998525a0fdc8907203658.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 108,
+ 79,
+ 502,
+ 115.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 108,
+ 115.0,
+ 502,
+ 151.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 108,
+ 151.0,
+ 502,
+ 187.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 197,
+ 506,
+ 237
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 197,
+ 505,
+ 209
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 197,
+ 505,
+ 209
+ ],
+ "score": 1.0,
+ "content": "Figure 9: (Best seen in pdf) (Left) Bidirectional prediction results on two-object pushing. More in Appendix",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 207,
+ 506,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 207,
+ 506,
+ 218
+ ],
+ "score": 1.0,
+ "content": "Fig 15. (Right) When used with a VAE (Sec 3.4), our approach captures residual stochasticity at the bottleneck.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 216,
+ 506,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 216,
+ 302,
+ 228
+ ],
+ "score": 1.0,
+ "content": "In these results from the pick-and-place task, GENMIN",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 303,
+ 219,
+ 309,
+ 226
+ ],
+ "score": 0.37,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 309,
+ 216,
+ 506,
+ 228
+ ],
+ "score": 1.0,
+ "content": "VAE produces images that are all of the arm in contact",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 226,
+ 474,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 226,
+ 474,
+ 239
+ ],
+ "score": 1.0,
+ "content": "with the object on the table, but at different points on the object, and with different arm/gripper poses.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5
+ }
+ ],
+ "index": 2.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 253,
+ 506,
+ 286
+ ],
+ "lines": [],
+ "index": 8,
+ "bbox_fs": [
+ 105,
+ 253,
+ 506,
+ 288
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 292,
+ 304,
+ 412
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 304,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 167,
+ 304
+ ],
+ "score": 1.0,
+ "content": "As indicated in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 167,
+ 292,
+ 197,
+ 303
+ ],
+ "score": 0.25,
+ "content": "{ \\mathrm { S e c } } 3 . 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 197,
+ 291,
+ 304,
+ 304
+ ],
+ "score": 1.0,
+ "content": "and Eq 3, TAP may also be",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 303,
+ 305,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 303,
+ 305,
+ 315
+ ],
+ "score": 1.0,
+ "content": "applied recursively. Fig 7 compares consecutive",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 314,
+ 305,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 314,
+ 305,
+ 326
+ ],
+ "score": 1.0,
+ "content": "subgoals for the pick-and-place task produced by",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 325,
+ 306,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 325,
+ 306,
+ 335
+ ],
+ "score": 1.0,
+ "content": "recursive TAP versus a recursive fixed-time model.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 335,
+ 306,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 335,
+ 172,
+ 348
+ ],
+ "score": 1.0,
+ "content": "Recursion level",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 172,
+ 336,
+ 198,
+ 346
+ ],
+ "score": 0.92,
+ "content": "r = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 198,
+ 335,
+ 306,
+ 348
+ ],
+ "score": 1.0,
+ "content": "refers to the first subgoal,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 347,
+ 305,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 347,
+ 124,
+ 358
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 124,
+ 347,
+ 150,
+ 357
+ ],
+ "score": 0.89,
+ "content": "r = 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 150,
+ 347,
+ 305,
+ 358
+ ],
+ "score": 1.0,
+ "content": "refers to the subgoal generated when",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 357,
+ 305,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 357,
+ 305,
+ 370
+ ],
+ "score": 1.0,
+ "content": "the first subgoal is provided as the goal input (start",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 369,
+ 305,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 369,
+ 305,
+ 381
+ ],
+ "score": 1.0,
+ "content": "input is unchanged). In example #1, FIX struggles",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 379,
+ 306,
+ 391
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 379,
+ 306,
+ 391
+ ],
+ "score": 1.0,
+ "content": "while “ours” identifies the “place” bottleneck at",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 107,
+ 390,
+ 306,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 391,
+ 131,
+ 401
+ ],
+ "score": 0.88,
+ "content": "r = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 132,
+ 390,
+ 306,
+ 402
+ ],
+ "score": 1.0,
+ "content": ", and subsequently the “pick” botleneck at",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 402,
+ 135,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 402,
+ 131,
+ 412
+ ],
+ "score": 0.84,
+ "content": "r = 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 131,
+ 402,
+ 135,
+ 412
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 15,
+ "bbox_fs": [
+ 105,
+ 291,
+ 306,
+ 412
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 312,
+ 291,
+ 504,
+ 371
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 312,
+ 291,
+ 504,
+ 371
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 312,
+ 291,
+ 504,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 312,
+ 291,
+ 504,
+ 371
+ ],
+ "score": 0.969,
+ "type": "image",
+ "image_path": "fe2aa8133a51dbd41e95dc77f0fc36f12af4aaa53458e7768be0fa33c485273a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 23.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 312,
+ 291,
+ 504,
+ 304.3333333333333
+ ],
+ "spans": [],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 312,
+ 304.3333333333333,
+ 504,
+ 317.66666666666663
+ ],
+ "spans": [],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 312,
+ 317.66666666666663,
+ 504,
+ 330.99999999999994
+ ],
+ "spans": [],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 312,
+ 330.99999999999994,
+ 504,
+ 344.33333333333326
+ ],
+ "spans": [],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 312,
+ 344.33333333333326,
+ 504,
+ 357.6666666666666
+ ],
+ "spans": [],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 312,
+ 357.6666666666666,
+ 504,
+ 370.9999999999999
+ ],
+ "spans": [],
+ "index": 26
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 312,
+ 376,
+ 505,
+ 397
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 312,
+ 376,
+ 505,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 312,
+ 376,
+ 505,
+ 387
+ ],
+ "score": 1.0,
+ "content": "Figure 7: Recursive bidirectional prediction on pick-",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 312,
+ 387,
+ 481,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 312,
+ 387,
+ 352,
+ 397
+ ],
+ "score": 1.0,
+ "content": "and-place.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 352,
+ 387,
+ 375,
+ 396
+ ],
+ "score": 0.88,
+ "content": "r = 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 375,
+ 387,
+ 454,
+ 397
+ ],
+ "score": 1.0,
+ "content": "is earlier in time than",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 455,
+ 387,
+ 477,
+ 396
+ ],
+ "score": 0.88,
+ "content": "r = 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 477,
+ 387,
+ 481,
+ 397
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27.5
+ }
+ ],
+ "index": 25.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 418,
+ 354,
+ 550
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 417,
+ 356,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 417,
+ 356,
+ 431
+ ],
+ "score": 1.0,
+ "content": "Finally, we test on “BAIR pushing” (Ebert et al., 2017), a",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 429,
+ 355,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 355,
+ 442
+ ],
+ "score": 1.0,
+ "content": "real-world dataset that is commonly used in visual prediction",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 440,
+ 355,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 440,
+ 355,
+ 452
+ ],
+ "score": 1.0,
+ "content": "tasks. The data consists of 30-frame clips of random motions",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 451,
+ 355,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 451,
+ 355,
+ 463
+ ],
+ "score": 1.0,
+ "content": "of a Sawyer arm tabletop. While this dataset does not have",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 462,
+ 355,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 296,
+ 474
+ ],
+ "score": 1.0,
+ "content": "natural bottlenecks like in grasping, TAP (min",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 296,
+ 462,
+ 306,
+ 473
+ ],
+ "score": 0.84,
+ "content": "\\ell _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 307,
+ 462,
+ 355,
+ 474
+ ],
+ "score": 1.0,
+ "content": "error 0.046",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 473,
+ 356,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 473,
+ 356,
+ 485
+ ],
+ "score": 1.0,
+ "content": "at match-step 15.42) still performs better than FIX (0.059 at",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 484,
+ 356,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 484,
+ 356,
+ 496
+ ],
+ "score": 1.0,
+ "content": "15.29). Qualitatively, as Fig 8 shows, even though BAIR push-",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 495,
+ 356,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 356,
+ 507
+ ],
+ "score": 1.0,
+ "content": "ing contains incoherent random arm motions, TAP consistently",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 506,
+ 356,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 506,
+ 356,
+ 518
+ ],
+ "score": 1.0,
+ "content": "produces predictions that plausibly lie on the path from start to",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 517,
+ 355,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 517,
+ 355,
+ 529
+ ],
+ "score": 1.0,
+ "content": "goal image. In example #1, given a start and goal image with",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 528,
+ 355,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 528,
+ 355,
+ 540
+ ],
+ "score": 1.0,
+ "content": "one object displaced, “ours” correctly moves the arm to the",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 538,
+ 309,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 538,
+ 309,
+ 552
+ ],
+ "score": 1.0,
+ "content": "object before displacement, whereas FIX struggles.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 34.5,
+ "bbox_fs": [
+ 105,
+ 417,
+ 356,
+ 552
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 362,
+ 418,
+ 504,
+ 495
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 362,
+ 418,
+ 504,
+ 495
+ ],
+ "group_id": 2,
+ "lines": [
+ {
+ "bbox": [
+ 362,
+ 418,
+ 504,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 362,
+ 418,
+ 504,
+ 495
+ ],
+ "score": 0.965,
+ "type": "image",
+ "image_path": "9385c328efd3ef9d40d2aebb6da428efee79f90d0d1504ea5d2217964752ce67.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 41.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 362,
+ 418,
+ 504,
+ 456.5
+ ],
+ "spans": [],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 362,
+ 456.5,
+ 504,
+ 495.0
+ ],
+ "spans": [],
+ "index": 42
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 362,
+ 502,
+ 505,
+ 542
+ ],
+ "group_id": 2,
+ "lines": [
+ {
+ "bbox": [
+ 361,
+ 501,
+ 506,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 361,
+ 501,
+ 506,
+ 513
+ ],
+ "score": 1.0,
+ "content": "Figure 8: Bidirectional prediction re-",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 362,
+ 512,
+ 505,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 362,
+ 512,
+ 505,
+ 523
+ ],
+ "score": 1.0,
+ "content": "sults on BAIR pushing data. The first",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 362,
+ 522,
+ 505,
+ 533
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 362,
+ 522,
+ 505,
+ 533
+ ],
+ "score": 1.0,
+ "content": "two columns are the inputs, and the next",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 361,
+ 532,
+ 495,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 361,
+ 532,
+ 495,
+ 542
+ ],
+ "score": 1.0,
+ "content": "two correspond to FIX and GENMIN.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 44.5
+ }
+ ],
+ "index": 43.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 567,
+ 396,
+ 687
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 567,
+ 396,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 567,
+ 396,
+ 579
+ ],
+ "score": 1.0,
+ "content": "Bottleneck discovery frequency. We have thus far relied on qualitative",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 578,
+ 396,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 578,
+ 396,
+ 590
+ ],
+ "score": 1.0,
+ "content": "results to assess how often our approach finds coherent bottlenecks. For",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 589,
+ 397,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 589,
+ 397,
+ 602
+ ],
+ "score": 1.0,
+ "content": "pushing, we test bottleneck discovery frequency more quantitatively. We",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 598,
+ 397,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 598,
+ 397,
+ 614
+ ],
+ "score": 1.0,
+ "content": "make the reasonable assumption that bottlenecks in 2-object pushing",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 106,
+ 611,
+ 398,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 611,
+ 398,
+ 623
+ ],
+ "score": 1.0,
+ "content": "correspond to states where one object is pushed and the other is in place.",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 621,
+ 397,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 397,
+ 634
+ ],
+ "score": 1.0,
+ "content": "Our metric exploits knowledge of true object positions at start and goal",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 106,
+ 633,
+ 396,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 396,
+ 644
+ ],
+ "score": 1.0,
+ "content": "states. First, for this evaluation, we restrict both GENMIN and FIX to",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 397,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 397,
+ 656
+ ],
+ "score": 1.0,
+ "content": "synthesize predictions purely by warping and masking inputs. Thus, we",
+ "type": "text"
+ }
+ ],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 105,
+ 655,
+ 397,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 655,
+ 397,
+ 667
+ ],
+ "score": 1.0,
+ "content": "can track where the pixels at ground truth object locations in start and",
+ "type": "text"
+ }
+ ],
+ "index": 57
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 397,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 397,
+ 678
+ ],
+ "score": 1.0,
+ "content": "goal images end up in the prediction, i.e., where did each object move?",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 396,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 396,
+ 689
+ ],
+ "score": 1.0,
+ "content": "We then compute a score that may be thresholded to detect when only",
+ "type": "text"
+ }
+ ],
+ "index": 60
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 701
+ ],
+ "score": 1.0,
+ "content": "one object moves (details in Appendix F). As an intuitive example, suppose that the two objects are",
+ "type": "text"
+ }
+ ],
+ "index": 62
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 159,
+ 711
+ ],
+ "score": 1.0,
+ "content": "displaced by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 160,
+ 699,
+ 186,
+ 709
+ ],
+ "score": 0.68,
+ "content": "1 0 \\mathrm { c m }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 186,
+ 699,
+ 203,
+ 711
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 204,
+ 699,
+ 230,
+ 709
+ ],
+ "score": 0.65,
+ "content": "1 5 \\mathrm { { c m } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 230,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "respectively between start and goal frames. Suppose further that our",
+ "type": "text"
+ }
+ ],
+ "index": 63
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "predictor predicts the start frame as its output. Then its distance score would be computed to be 10",
+ "type": "text"
+ }
+ ],
+ "index": 64
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 507,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 256,
+ 733
+ ],
+ "score": 1.0,
+ "content": "cm. For all values of threshold below",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 257,
+ 721,
+ 282,
+ 731
+ ],
+ "score": 0.67,
+ "content": "1 0 \\mathrm { c m }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 282,
+ 720,
+ 507,
+ 733
+ ],
+ "score": 1.0,
+ "content": ", this would be counted as a bottleneck discovery failure.",
+ "type": "text"
+ }
+ ],
+ "index": 65
+ }
+ ],
+ "index": 53,
+ "bbox_fs": [
+ 105,
+ 567,
+ 398,
+ 689
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 406,
+ 570,
+ 500,
+ 647
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 406,
+ 570,
+ 500,
+ 647
+ ],
+ "group_id": 3,
+ "lines": [
+ {
+ "bbox": [
+ 406,
+ 570,
+ 500,
+ 647
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 406,
+ 570,
+ 500,
+ 647
+ ],
+ "score": 0.958,
+ "type": "image",
+ "image_path": "c98be0a524e6a6ba7b94ce1b7a36d75a2305e1440f91fbf772d0f3efa5ba64f2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 53.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 406,
+ 570,
+ 500,
+ 608.5
+ ],
+ "spans": [],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 406,
+ 608.5,
+ 500,
+ 647.0
+ ],
+ "spans": [],
+ "index": 56
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 403,
+ 659,
+ 505,
+ 679
+ ],
+ "group_id": 3,
+ "lines": [
+ {
+ "bbox": [
+ 403,
+ 659,
+ 506,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 403,
+ 659,
+ 506,
+ 669
+ ],
+ "score": 1.0,
+ "content": "Figure 10: Bottleneck fre-",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 403,
+ 670,
+ 501,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 403,
+ 670,
+ 501,
+ 679
+ ],
+ "score": 1.0,
+ "content": "quency vs. score threshold",
+ "type": "text"
+ }
+ ],
+ "index": 61
+ }
+ ],
+ "index": 59.5
+ }
+ ],
+ "index": 56.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 689,
+ 505,
+ 732
+ ],
+ "lines": [],
+ "index": 63.5,
+ "bbox_fs": [
+ 105,
+ 687,
+ 507,
+ 733
+ ],
+ "lines_deleted": true
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 82,
+ 504,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 390,
+ 95
+ ],
+ "score": 1.0,
+ "content": "As Fig 10 shows, GENMIN predicts bottlenecks much more frequently",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 391,
+ 83,
+ 419,
+ 93
+ ],
+ "score": 0.88,
+ "content": "\\sim 6 0 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 82,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": "of the time) than FIX.",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 107
+ ],
+ "score": 1.0,
+ "content": "As hypothesized, our time-agnostic approach does indeed identify and latch on to low-uncertainty",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 238,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 238,
+ 116
+ ],
+ "score": 1.0,
+ "content": "states to improve its predictions.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 139,
+ 375,
+ 228
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 139,
+ 376,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 139,
+ 376,
+ 151
+ ],
+ "score": 1.0,
+ "content": "Hierarchical planning evaluation. Finally, we discuss experi-",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 150,
+ 376,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 150,
+ 376,
+ 162
+ ],
+ "score": 1.0,
+ "content": "ments directly evaluating our intermediate predictions as visual",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 161,
+ 376,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 161,
+ 376,
+ 173
+ ],
+ "score": 1.0,
+ "content": "subgoals for hierarchical planning for pushing tasks. A forward",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 172,
+ 376,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 172,
+ 376,
+ 184
+ ],
+ "score": 1.0,
+ "content": "Visual MPC planner (Ebert et al., 2017) accepts the subgoal object",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 183,
+ 376,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 183,
+ 376,
+ 195
+ ],
+ "score": 1.0,
+ "content": "positions (computed as above for evaluating bottlenecks). Start and",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 194,
+ 376,
+ 206
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 194,
+ 376,
+ 206
+ ],
+ "score": 1.0,
+ "content": "goal object positions are also known. Internally, Visual MPC makes",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 205,
+ 376,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 205,
+ 376,
+ 217
+ ],
+ "score": 1.0,
+ "content": "action-conditioned fixed-time forward predictions of future object",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 216,
+ 376,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 216,
+ 376,
+ 228
+ ],
+ "score": 1.0,
+ "content": "positions to find an action sequence that reaches the subgoal object",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 6.5
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 385,
+ 141,
+ 504,
+ 189
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 385,
+ 141,
+ 504,
+ 189
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 385,
+ 141,
+ 504,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 385,
+ 141,
+ 504,
+ 189
+ ],
+ "score": 0.971,
+ "html": " | 2-object | 3-object |
| direct | 12.9±0.6 | 15.8±0.6 |
| FIX | 12.5±0.5 | 17.6±0.6 |
| GENMIN | 11.9±0.6 | 12.9±0.7 |
",
+ "type": "table",
+ "image_path": "91aa2871f959f5c4caab3f23c69cc737c597c3a5f2593fa436f4e4dcd82aec90.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 11.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 385,
+ 141,
+ 504,
+ 165.0
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 385,
+ 165.0,
+ 504,
+ 189.0
+ ],
+ "spans": [],
+ "index": 12
+ }
+ ]
+ }
+ ],
+ "index": 11.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 384,
+ 197,
+ 505,
+ 217
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 382,
+ 196,
+ 506,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 382,
+ 196,
+ 506,
+ 208
+ ],
+ "score": 1.0,
+ "content": "Table 2: Multi-object pushing er-",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 383,
+ 207,
+ 430,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 383,
+ 207,
+ 430,
+ 218
+ ],
+ "score": 1.0,
+ "content": "rors (in cm).",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 110,
+ 227,
+ 487,
+ 238
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 225,
+ 489,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 225,
+ 489,
+ 240
+ ],
+ "score": 1.0,
+ "content": "positions, with a planning horizon of 15. Additional implementation details are in Appendix G.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 244,
+ 505,
+ 354
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 243,
+ 505,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 243,
+ 505,
+ 256
+ ],
+ "score": 1.0,
+ "content": "Given start and goal images, our model produces a subgoal. Visual MPC plans towards this subgoal",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 254,
+ 506,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 254,
+ 506,
+ 268
+ ],
+ "score": 1.0,
+ "content": "for half the episode length, then switches to the final goal. We compare this scheme against (i)",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 265,
+ 505,
+ 278
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 265,
+ 505,
+ 278
+ ],
+ "score": 1.0,
+ "content": "“direct”: planning directly towards the final goal for the entire episode length, and (ii) FIX: subgoals",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 276,
+ 506,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 276,
+ 506,
+ 289
+ ],
+ "score": 1.0,
+ "content": "from a center-frame predictor. The error measure is the mean of object distances to goal states (lower",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 288,
+ 506,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 288,
+ 407,
+ 300
+ ],
+ "score": 1.0,
+ "content": "is better). As an upper bound, single-object pushing with the planner yields",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 408,
+ 288,
+ 434,
+ 298
+ ],
+ "score": 0.85,
+ "content": "{ \\sim } 5 \\mathrm { c m }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 434,
+ 288,
+ 506,
+ 300
+ ],
+ "score": 1.0,
+ "content": "error. Results for",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 299,
+ 505,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 299,
+ 505,
+ 312
+ ],
+ "score": 1.0,
+ "content": "two-object and three-object pushing are shown in Table 2. GENMIN does best on both, but especially",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 309,
+ 506,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 309,
+ 506,
+ 322
+ ],
+ "score": 1.0,
+ "content": "on the more complex three-object task. Since Visual MPC has thus far been demonstrated to work",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 321,
+ 506,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 321,
+ 506,
+ 333
+ ],
+ "score": 1.0,
+ "content": "only on pushing tasks, our hierarchical planning evaluation is also limited to this task. Going forward,",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 331,
+ 505,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 331,
+ 505,
+ 344
+ ],
+ "score": 1.0,
+ "content": "we plan to adapt Visual MPC to allow testing TAP on more complex temporally extended tasks like",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 342,
+ 444,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 342,
+ 444,
+ 355
+ ],
+ "score": 1.0,
+ "content": "block-stacking, where direct planning breaks down and subgoals offer greater value.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 20.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 381,
+ 201,
+ 393
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 378,
+ 203,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 378,
+ 203,
+ 397
+ ],
+ "score": 1.0,
+ "content": "5 CONCLUSIONS",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 412,
+ 505,
+ 522
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 412,
+ 506,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 412,
+ 506,
+ 425
+ ],
+ "score": 1.0,
+ "content": "The standard paradigm for prediction tasks demands that a predictor not only make good predictions,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 424,
+ 505,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 424,
+ 505,
+ 437
+ ],
+ "score": 1.0,
+ "content": "but that it make them on a set schedule. We have argued for redefining the prediction task so that the",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 435,
+ 505,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 435,
+ 505,
+ 447
+ ],
+ "score": 1.0,
+ "content": "predictor need only care that its prediction occur at some time, rather than that it occur at a specific",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 459
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 459
+ ],
+ "score": 1.0,
+ "content": "scheduled time. We define this time-agnostic prediction task and propose novel technical approaches",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 457,
+ 506,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 457,
+ 506,
+ 469
+ ],
+ "score": 1.0,
+ "content": "to solve it, that require relatively small changes to standard prediction methods. Our results show that",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 468,
+ 505,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 468,
+ 505,
+ 480
+ ],
+ "score": 1.0,
+ "content": "reframing prediction objectives in this way yields higher quality predictions that are also semantically",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 479,
+ 504,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 479,
+ 504,
+ 491
+ ],
+ "score": 1.0,
+ "content": "coherent—unattached to a rigid schedule of regularly specified timestamps, model predictions instead",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 489,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 489,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "naturally attach to specific semantic “bottleneck” events, like a grasp. In our preliminary experiments",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 513
+ ],
+ "score": 1.0,
+ "content": "with a hierarchical visual planner, our results suggest that such predictions could serve as useful",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 511,
+ 219,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 511,
+ 219,
+ 524
+ ],
+ "score": 1.0,
+ "content": "subgoals for complex tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 31.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 528,
+ 505,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 529,
+ 504,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 529,
+ 504,
+ 540
+ ],
+ "score": 1.0,
+ "content": "In future work, we would like to address some limitations of our TAP formulation, of which we will",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 539,
+ 505,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 539,
+ 505,
+ 551
+ ],
+ "score": 1.0,
+ "content": "mention two here. First, TAP currently benefits not only from selecting which times to predict, but",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "score": 1.0,
+ "content": "also from not having to provide timestamps attached to its predictions. We would like to study: could",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 575
+ ],
+ "score": 1.0,
+ "content": "we retain the benefits of time-agnostic prediction while also providing timestamps for when each",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 572,
+ 504,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 504,
+ 584
+ ],
+ "score": 1.0,
+ "content": "predicted state will occur? Second, our current TAP formulation may not generalize to prediction",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 583,
+ 506,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 583,
+ 506,
+ 596
+ ],
+ "score": 1.0,
+ "content": "problems in all settings of interest. As an example, for videos of juggling or waving, which involve",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 606
+ ],
+ "score": 1.0,
+ "content": "repeated frames, TAP might collapse to predicting the input state repeatedly. We would like to",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 604,
+ 505,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 411,
+ 619
+ ],
+ "score": 1.0,
+ "content": "investigate more general TAP formulations: for example, rather than choosing",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 412,
+ 605,
+ 422,
+ 617
+ ],
+ "score": 0.89,
+ "content": "\\mathcal { E } _ { t } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 422,
+ 604,
+ 505,
+ 619
+ ],
+ "score": 1.0,
+ "content": "in Eq 5 to encourage",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 629
+ ],
+ "score": 1.0,
+ "content": "predicting farther away times, we could conceivably penalize predictions that look too similar to the",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "input context frames. More broadly, we believe that our results thus far hold great promise for many",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 507,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 507,
+ 652
+ ],
+ "score": 1.0,
+ "content": "applications of prediction including hierarchical planning and model-based reinforcement learning,",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 289,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 289,
+ 661
+ ],
+ "score": 1.0,
+ "content": "and we hope to build further on these results.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 42.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 666,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 677
+ ],
+ "score": 1.0,
+ "content": "Acknowledgements. We thank Alex Lee and Chelsea Finn for helpful discussions and Sudeep Dasari",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "for help with the simulation framework and for generating the simulated data used in this work. We",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "thank Kate Rakelly, Kyle Hsu, and Allan Jabri for feedback on early drafts. This work was supported",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "by Berkeley DeepDrive, NSF IIS-1614653, NSF IIS-1633310, and an Office of Naval Research",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 708,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 506,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Young Investigator Program award. The NVIDIA DGX-1 used for this research was donated by the",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 198,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 198,
+ 734
+ ],
+ "score": 1.0,
+ "content": "NVIDIA Corporation.",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ }
+ ],
+ "index": 51.5
+ }
+ ],
+ "page_idx": 8,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "9",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 82,
+ 504,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 390,
+ 95
+ ],
+ "score": 1.0,
+ "content": "As Fig 10 shows, GENMIN predicts bottlenecks much more frequently",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 391,
+ 83,
+ 419,
+ 93
+ ],
+ "score": 0.88,
+ "content": "\\sim 6 0 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 82,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": "of the time) than FIX.",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 107
+ ],
+ "score": 1.0,
+ "content": "As hypothesized, our time-agnostic approach does indeed identify and latch on to low-uncertainty",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 238,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 238,
+ 116
+ ],
+ "score": 1.0,
+ "content": "states to improve its predictions.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1,
+ "bbox_fs": [
+ 105,
+ 82,
+ 506,
+ 116
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 139,
+ 375,
+ 228
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 139,
+ 376,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 139,
+ 376,
+ 151
+ ],
+ "score": 1.0,
+ "content": "Hierarchical planning evaluation. Finally, we discuss experi-",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 150,
+ 376,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 150,
+ 376,
+ 162
+ ],
+ "score": 1.0,
+ "content": "ments directly evaluating our intermediate predictions as visual",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 161,
+ 376,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 161,
+ 376,
+ 173
+ ],
+ "score": 1.0,
+ "content": "subgoals for hierarchical planning for pushing tasks. A forward",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 172,
+ 376,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 172,
+ 376,
+ 184
+ ],
+ "score": 1.0,
+ "content": "Visual MPC planner (Ebert et al., 2017) accepts the subgoal object",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 183,
+ 376,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 183,
+ 376,
+ 195
+ ],
+ "score": 1.0,
+ "content": "positions (computed as above for evaluating bottlenecks). Start and",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 194,
+ 376,
+ 206
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 194,
+ 376,
+ 206
+ ],
+ "score": 1.0,
+ "content": "goal object positions are also known. Internally, Visual MPC makes",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 205,
+ 376,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 205,
+ 376,
+ 217
+ ],
+ "score": 1.0,
+ "content": "action-conditioned fixed-time forward predictions of future object",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 216,
+ 376,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 216,
+ 376,
+ 228
+ ],
+ "score": 1.0,
+ "content": "positions to find an action sequence that reaches the subgoal object",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 6.5,
+ "bbox_fs": [
+ 105,
+ 139,
+ 376,
+ 228
+ ]
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 385,
+ 141,
+ 504,
+ 189
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 385,
+ 141,
+ 504,
+ 189
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 385,
+ 141,
+ 504,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 385,
+ 141,
+ 504,
+ 189
+ ],
+ "score": 0.971,
+ "html": " | 2-object | 3-object |
| direct | 12.9±0.6 | 15.8±0.6 |
| FIX | 12.5±0.5 | 17.6±0.6 |
| GENMIN | 11.9±0.6 | 12.9±0.7 |
",
+ "type": "table",
+ "image_path": "91aa2871f959f5c4caab3f23c69cc737c597c3a5f2593fa436f4e4dcd82aec90.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 11.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 385,
+ 141,
+ 504,
+ 165.0
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 385,
+ 165.0,
+ 504,
+ 189.0
+ ],
+ "spans": [],
+ "index": 12
+ }
+ ]
+ }
+ ],
+ "index": 11.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 384,
+ 197,
+ 505,
+ 217
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 382,
+ 196,
+ 506,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 382,
+ 196,
+ 506,
+ 208
+ ],
+ "score": 1.0,
+ "content": "Table 2: Multi-object pushing er-",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 383,
+ 207,
+ 430,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 383,
+ 207,
+ 430,
+ 218
+ ],
+ "score": 1.0,
+ "content": "rors (in cm).",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5,
+ "bbox_fs": [
+ 382,
+ 196,
+ 506,
+ 218
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 110,
+ 227,
+ 487,
+ 238
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 225,
+ 489,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 225,
+ 489,
+ 240
+ ],
+ "score": 1.0,
+ "content": "positions, with a planning horizon of 15. Additional implementation details are in Appendix G.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 15,
+ "bbox_fs": [
+ 108,
+ 225,
+ 489,
+ 240
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 244,
+ 505,
+ 354
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 243,
+ 505,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 243,
+ 505,
+ 256
+ ],
+ "score": 1.0,
+ "content": "Given start and goal images, our model produces a subgoal. Visual MPC plans towards this subgoal",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 254,
+ 506,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 254,
+ 506,
+ 268
+ ],
+ "score": 1.0,
+ "content": "for half the episode length, then switches to the final goal. We compare this scheme against (i)",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 265,
+ 505,
+ 278
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 265,
+ 505,
+ 278
+ ],
+ "score": 1.0,
+ "content": "“direct”: planning directly towards the final goal for the entire episode length, and (ii) FIX: subgoals",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 276,
+ 506,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 276,
+ 506,
+ 289
+ ],
+ "score": 1.0,
+ "content": "from a center-frame predictor. The error measure is the mean of object distances to goal states (lower",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 288,
+ 506,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 288,
+ 407,
+ 300
+ ],
+ "score": 1.0,
+ "content": "is better). As an upper bound, single-object pushing with the planner yields",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 408,
+ 288,
+ 434,
+ 298
+ ],
+ "score": 0.85,
+ "content": "{ \\sim } 5 \\mathrm { c m }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 434,
+ 288,
+ 506,
+ 300
+ ],
+ "score": 1.0,
+ "content": "error. Results for",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 299,
+ 505,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 299,
+ 505,
+ 312
+ ],
+ "score": 1.0,
+ "content": "two-object and three-object pushing are shown in Table 2. GENMIN does best on both, but especially",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 309,
+ 506,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 309,
+ 506,
+ 322
+ ],
+ "score": 1.0,
+ "content": "on the more complex three-object task. Since Visual MPC has thus far been demonstrated to work",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 321,
+ 506,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 321,
+ 506,
+ 333
+ ],
+ "score": 1.0,
+ "content": "only on pushing tasks, our hierarchical planning evaluation is also limited to this task. Going forward,",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 331,
+ 505,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 331,
+ 505,
+ 344
+ ],
+ "score": 1.0,
+ "content": "we plan to adapt Visual MPC to allow testing TAP on more complex temporally extended tasks like",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 342,
+ 444,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 342,
+ 444,
+ 355
+ ],
+ "score": 1.0,
+ "content": "block-stacking, where direct planning breaks down and subgoals offer greater value.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 20.5,
+ "bbox_fs": [
+ 105,
+ 243,
+ 506,
+ 355
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 381,
+ 201,
+ 393
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 378,
+ 203,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 378,
+ 203,
+ 397
+ ],
+ "score": 1.0,
+ "content": "5 CONCLUSIONS",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 412,
+ 505,
+ 522
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 412,
+ 506,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 412,
+ 506,
+ 425
+ ],
+ "score": 1.0,
+ "content": "The standard paradigm for prediction tasks demands that a predictor not only make good predictions,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 424,
+ 505,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 424,
+ 505,
+ 437
+ ],
+ "score": 1.0,
+ "content": "but that it make them on a set schedule. We have argued for redefining the prediction task so that the",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 435,
+ 505,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 435,
+ 505,
+ 447
+ ],
+ "score": 1.0,
+ "content": "predictor need only care that its prediction occur at some time, rather than that it occur at a specific",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 459
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 459
+ ],
+ "score": 1.0,
+ "content": "scheduled time. We define this time-agnostic prediction task and propose novel technical approaches",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 457,
+ 506,
+ 469
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 457,
+ 506,
+ 469
+ ],
+ "score": 1.0,
+ "content": "to solve it, that require relatively small changes to standard prediction methods. Our results show that",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 468,
+ 505,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 468,
+ 505,
+ 480
+ ],
+ "score": 1.0,
+ "content": "reframing prediction objectives in this way yields higher quality predictions that are also semantically",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 479,
+ 504,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 479,
+ 504,
+ 491
+ ],
+ "score": 1.0,
+ "content": "coherent—unattached to a rigid schedule of regularly specified timestamps, model predictions instead",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 489,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 489,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "naturally attach to specific semantic “bottleneck” events, like a grasp. In our preliminary experiments",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 513
+ ],
+ "score": 1.0,
+ "content": "with a hierarchical visual planner, our results suggest that such predictions could serve as useful",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 511,
+ 219,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 511,
+ 219,
+ 524
+ ],
+ "score": 1.0,
+ "content": "subgoals for complex tasks.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 31.5,
+ "bbox_fs": [
+ 105,
+ 412,
+ 506,
+ 524
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 528,
+ 505,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 529,
+ 504,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 529,
+ 504,
+ 540
+ ],
+ "score": 1.0,
+ "content": "In future work, we would like to address some limitations of our TAP formulation, of which we will",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 539,
+ 505,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 539,
+ 505,
+ 551
+ ],
+ "score": 1.0,
+ "content": "mention two here. First, TAP currently benefits not only from selecting which times to predict, but",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 563
+ ],
+ "score": 1.0,
+ "content": "also from not having to provide timestamps attached to its predictions. We would like to study: could",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 505,
+ 575
+ ],
+ "score": 1.0,
+ "content": "we retain the benefits of time-agnostic prediction while also providing timestamps for when each",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 572,
+ 504,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 504,
+ 584
+ ],
+ "score": 1.0,
+ "content": "predicted state will occur? Second, our current TAP formulation may not generalize to prediction",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 583,
+ 506,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 583,
+ 506,
+ 596
+ ],
+ "score": 1.0,
+ "content": "problems in all settings of interest. As an example, for videos of juggling or waving, which involve",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 606
+ ],
+ "score": 1.0,
+ "content": "repeated frames, TAP might collapse to predicting the input state repeatedly. We would like to",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 604,
+ 505,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 411,
+ 619
+ ],
+ "score": 1.0,
+ "content": "investigate more general TAP formulations: for example, rather than choosing",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 412,
+ 605,
+ 422,
+ 617
+ ],
+ "score": 0.89,
+ "content": "\\mathcal { E } _ { t } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 422,
+ 604,
+ 505,
+ 619
+ ],
+ "score": 1.0,
+ "content": "in Eq 5 to encourage",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 629
+ ],
+ "score": 1.0,
+ "content": "predicting farther away times, we could conceivably penalize predictions that look too similar to the",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "input context frames. More broadly, we believe that our results thus far hold great promise for many",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 507,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 507,
+ 652
+ ],
+ "score": 1.0,
+ "content": "applications of prediction including hierarchical planning and model-based reinforcement learning,",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 289,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 289,
+ 661
+ ],
+ "score": 1.0,
+ "content": "and we hope to build further on these results.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 42.5,
+ "bbox_fs": [
+ 105,
+ 529,
+ 507,
+ 661
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 666,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 677
+ ],
+ "score": 1.0,
+ "content": "Acknowledgements. We thank Alex Lee and Chelsea Finn for helpful discussions and Sudeep Dasari",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "for help with the simulation framework and for generating the simulated data used in this work. We",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "thank Kate Rakelly, Kyle Hsu, and Allan Jabri for feedback on early drafts. This work was supported",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "by Berkeley DeepDrive, NSF IIS-1614653, NSF IIS-1633310, and an Office of Naval Research",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 105,
+ 708,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 506,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Young Investigator Program award. The NVIDIA DGX-1 used for this research was donated by the",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 198,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 198,
+ 734
+ ],
+ "score": 1.0,
+ "content": "NVIDIA Corporation.",
+ "type": "text"
+ }
+ ],
+ "index": 54
+ }
+ ],
+ "index": 51.5,
+ "bbox_fs": [
+ 105,
+ 666,
+ 506,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 81,
+ 176,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 176,
+ 94
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 176,
+ 94
+ ],
+ "score": 1.0,
+ "content": "REFERENCES",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 100,
+ 504,
+ 123
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 99,
+ 506,
+ 113
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 99,
+ 506,
+ 113
+ ],
+ "score": 1.0,
+ "content": "Mohammad Babaeizadeh, Chelsea Finn, Dumitru Erhan, Roy H Campbell, and Sergey Levine.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 117,
+ 111,
+ 327,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 111,
+ 327,
+ 123
+ ],
+ "score": 1.0,
+ "content": "Stochastic variational video prediction. ICLR, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 128,
+ 504,
+ 152
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 505,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 505,
+ 142
+ ],
+ "score": 1.0,
+ "content": "Pierre-Luc Bacon. On the bottleneck concept for options discovery. M.S. Thesis, McGill University,",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 115,
+ 139,
+ 143,
+ 153
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 139,
+ 143,
+ 153
+ ],
+ "score": 1.0,
+ "content": "2013.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 158,
+ 504,
+ 181
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 158,
+ 505,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 158,
+ 505,
+ 171
+ ],
+ "score": 1.0,
+ "content": "Moshe Bar. The proactive brain: memory for predictions. Philosophical Transactions of the Royal",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 115,
+ 169,
+ 361,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 169,
+ 361,
+ 182
+ ],
+ "score": 1.0,
+ "content": "Society B: Biological Sciences, 364(1521):1235–1243, 2009.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 187,
+ 504,
+ 211
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 186,
+ 505,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 186,
+ 505,
+ 201
+ ],
+ "score": 1.0,
+ "content": "Andy Clark. Whatever next? predictive brains, situated agents, and the future of cognitive science.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 116,
+ 199,
+ 331,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 199,
+ 331,
+ 210
+ ],
+ "score": 1.0,
+ "content": "Behavioral and brain sciences, 36(3):181–204, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 217,
+ 505,
+ 240
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 216,
+ 505,
+ 231
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 216,
+ 505,
+ 231
+ ],
+ "score": 1.0,
+ "content": "Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 115,
+ 228,
+ 219,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 228,
+ 219,
+ 240
+ ],
+ "score": 1.0,
+ "content": "arXiv:1802.07687, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 9.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 246,
+ 505,
+ 270
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 246,
+ 505,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 246,
+ 505,
+ 259
+ ],
+ "score": 1.0,
+ "content": "Frederik Ebert, Chelsea Finn, Alex X. Lee, and Sergey Levine. Self-supervised visual planning with",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 116,
+ 258,
+ 283,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 258,
+ 283,
+ 269
+ ],
+ "score": 1.0,
+ "content": "temporal skip connections. CORL, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 275,
+ 504,
+ 299
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 275,
+ 506,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 275,
+ 506,
+ 290
+ ],
+ "score": 1.0,
+ "content": "Chelsea Finn and Sergey Levine. Deep visual foresight for planning robot motion. In International",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 117,
+ 288,
+ 339,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 288,
+ 339,
+ 299
+ ],
+ "score": 1.0,
+ "content": "Conference on Robotics and Automation (ICRA), 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 304,
+ 505,
+ 329
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 305,
+ 505,
+ 318
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 305,
+ 505,
+ 318
+ ],
+ "score": 1.0,
+ "content": "Chelsea Finn, Ian Goodfellow, and Sergey Levine. Unsupervised learning for physical interaction",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 116,
+ 316,
+ 450,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 316,
+ 450,
+ 329
+ ],
+ "score": 1.0,
+ "content": "through video prediction. In Neural Information Processing Systems (NIPS), 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 334,
+ 504,
+ 369
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 348
+ ],
+ "score": 1.0,
+ "content": "Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 115,
+ 346,
+ 506,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 346,
+ 506,
+ 358
+ ],
+ "score": 1.0,
+ "content": "Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural informa-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 116,
+ 357,
+ 305,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 357,
+ 305,
+ 369
+ ],
+ "score": 1.0,
+ "content": "tion processing systems, pp. 2672–2680, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 375,
+ 504,
+ 409
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 375,
+ 505,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 375,
+ 505,
+ 388
+ ],
+ "score": 1.0,
+ "content": "Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 115,
+ 386,
+ 506,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 386,
+ 506,
+ 399
+ ],
+ "score": 1.0,
+ "content": "mapping. In Computer vision and pattern recognition, 2006 IEEE computer society conference on,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 116,
+ 398,
+ 275,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 398,
+ 275,
+ 410
+ ],
+ "score": 1.0,
+ "content": "volume 2, pp. 1735–1742. IEEE, 2006.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 415,
+ 378,
+ 428
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 415,
+ 379,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 415,
+ 379,
+ 430
+ ],
+ "score": 1.0,
+ "content": "Jakob Hohwy. The predictive mind. Oxford University Press, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 433,
+ 505,
+ 457
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 433,
+ 505,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 433,
+ 505,
+ 447
+ ],
+ "score": 1.0,
+ "content": "Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 116,
+ 445,
+ 307,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 445,
+ 307,
+ 456
+ ],
+ "score": 1.0,
+ "content": "conditional adversarial networks. CVPR, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 463,
+ 503,
+ 487
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 464,
+ 505,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 464,
+ 505,
+ 477
+ ],
+ "score": 1.0,
+ "content": "Dinesh Jayaraman and Kristen Grauman. Learning image representations tied to ego-motion. In",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 115,
+ 474,
+ 494,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 474,
+ 494,
+ 487
+ ],
+ "score": 1.0,
+ "content": "Proceedings of the IEEE International Conference on Computer Vision, pp. 1413–1421, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 492,
+ 504,
+ 516
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 506,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 506,
+ 506
+ ],
+ "score": 1.0,
+ "content": "Nal Kalchbrenner, Aaron van den Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyals, Alex Graves,",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 117,
+ 504,
+ 473,
+ 517
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 504,
+ 473,
+ 517
+ ],
+ "score": 1.0,
+ "content": "and Koray Kavukcuoglu. Video pixel networks. arXiv preprint arXiv:1610.00527, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 522,
+ 504,
+ 545
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 521,
+ 506,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 521,
+ 506,
+ 536
+ ],
+ "score": 1.0,
+ "content": "Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 115,
+ 533,
+ 214,
+ 544
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 533,
+ 214,
+ 544
+ ],
+ "score": 1.0,
+ "content": "arXiv:1312.6114, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 551,
+ 504,
+ 575
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "Kris M Kitani, Brian D Ziebart, James Andrew Bagnell, and Martial Hebert. Activity forecasting. In",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 116,
+ 562,
+ 409,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 562,
+ 409,
+ 576
+ ],
+ "score": 1.0,
+ "content": "European Conference on Computer Vision, pp. 201–214. Springer, 2012.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 580,
+ 506,
+ 615
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 579,
+ 506,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 579,
+ 506,
+ 594
+ ],
+ "score": 1.0,
+ "content": "Dirk P Kroese, Reuven Y Rubinstein, Izack Cohen, Sergey Porotsky, and Thomas Taimre. Cross-",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 592,
+ 507,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 592,
+ 507,
+ 604
+ ],
+ "score": 1.0,
+ "content": "entropy method. In Encyclopedia of Operations Research and Management Science, pp. 326–333.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 115,
+ 602,
+ 181,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 602,
+ 181,
+ 617
+ ],
+ "score": 1.0,
+ "content": "Springer, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 621,
+ 505,
+ 645
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 621,
+ 506,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 621,
+ 506,
+ 634
+ ],
+ "score": 1.0,
+ "content": "Anders Boesen Lindbo Larsen, Søren Kaae Sønderby, Hugo Larochelle, and Ole Winther. Autoen-",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 116,
+ 632,
+ 393,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 632,
+ 393,
+ 645
+ ],
+ "score": 1.0,
+ "content": "coding beyond pixels using a learned similarity metric. ICML, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 37.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 650,
+ 503,
+ 674
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 504,
+ 664
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 504,
+ 664
+ ],
+ "score": 1.0,
+ "content": "Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 117,
+ 662,
+ 439,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 662,
+ 439,
+ 673
+ ],
+ "score": 1.0,
+ "content": "Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 39.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 679,
+ 503,
+ 703
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 679,
+ 505,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 679,
+ 505,
+ 693
+ ],
+ "score": 1.0,
+ "content": "Michael Mathieu, Camille Couprie, and Yann LeCun. Deep multi-scale video prediction beyond",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 116,
+ 692,
+ 356,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 692,
+ 356,
+ 702
+ ],
+ "score": 1.0,
+ "content": "mean square error. arXiv preprint arXiv:1511.05440, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 41.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 709,
+ 503,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 708,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 708,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Amy McGovern and Andrew G Barto. Automatic discovery of subgoals in reinforcement learning",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 115,
+ 721,
+ 271,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 721,
+ 271,
+ 732
+ ],
+ "score": 1.0,
+ "content": "using diverse density. In ICML, 2001.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43.5
+ }
+ ],
+ "page_idx": 9,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 298,
+ 750,
+ 312,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 298,
+ 750,
+ 312,
+ 764
+ ],
+ "score": 1.0,
+ "content": "10",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 81,
+ 176,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 176,
+ 94
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 176,
+ 94
+ ],
+ "score": 1.0,
+ "content": "REFERENCES",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 100,
+ 504,
+ 123
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 99,
+ 506,
+ 113
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 99,
+ 506,
+ 113
+ ],
+ "score": 1.0,
+ "content": "Mohammad Babaeizadeh, Chelsea Finn, Dumitru Erhan, Roy H Campbell, and Sergey Levine.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 117,
+ 111,
+ 327,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 111,
+ 327,
+ 123
+ ],
+ "score": 1.0,
+ "content": "Stochastic variational video prediction. ICLR, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1.5,
+ "bbox_fs": [
+ 105,
+ 99,
+ 506,
+ 123
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 128,
+ 504,
+ 152
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 505,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 505,
+ 142
+ ],
+ "score": 1.0,
+ "content": "Pierre-Luc Bacon. On the bottleneck concept for options discovery. M.S. Thesis, McGill University,",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 115,
+ 139,
+ 143,
+ 153
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 139,
+ 143,
+ 153
+ ],
+ "score": 1.0,
+ "content": "2013.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5,
+ "bbox_fs": [
+ 105,
+ 128,
+ 505,
+ 153
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 158,
+ 504,
+ 181
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 158,
+ 505,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 158,
+ 505,
+ 171
+ ],
+ "score": 1.0,
+ "content": "Moshe Bar. The proactive brain: memory for predictions. Philosophical Transactions of the Royal",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 115,
+ 169,
+ 361,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 169,
+ 361,
+ 182
+ ],
+ "score": 1.0,
+ "content": "Society B: Biological Sciences, 364(1521):1235–1243, 2009.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5,
+ "bbox_fs": [
+ 105,
+ 158,
+ 505,
+ 182
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 187,
+ 504,
+ 211
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 186,
+ 505,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 186,
+ 505,
+ 201
+ ],
+ "score": 1.0,
+ "content": "Andy Clark. Whatever next? predictive brains, situated agents, and the future of cognitive science.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 116,
+ 199,
+ 331,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 199,
+ 331,
+ 210
+ ],
+ "score": 1.0,
+ "content": "Behavioral and brain sciences, 36(3):181–204, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5,
+ "bbox_fs": [
+ 106,
+ 186,
+ 505,
+ 210
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 217,
+ 505,
+ 240
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 216,
+ 505,
+ 231
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 216,
+ 505,
+ 231
+ ],
+ "score": 1.0,
+ "content": "Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 115,
+ 228,
+ 219,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 228,
+ 219,
+ 240
+ ],
+ "score": 1.0,
+ "content": "arXiv:1802.07687, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 9.5,
+ "bbox_fs": [
+ 105,
+ 216,
+ 505,
+ 240
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 246,
+ 505,
+ 270
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 246,
+ 505,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 246,
+ 505,
+ 259
+ ],
+ "score": 1.0,
+ "content": "Frederik Ebert, Chelsea Finn, Alex X. Lee, and Sergey Levine. Self-supervised visual planning with",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 116,
+ 258,
+ 283,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 258,
+ 283,
+ 269
+ ],
+ "score": 1.0,
+ "content": "temporal skip connections. CORL, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11.5,
+ "bbox_fs": [
+ 106,
+ 246,
+ 505,
+ 269
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 275,
+ 504,
+ 299
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 275,
+ 506,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 275,
+ 506,
+ 290
+ ],
+ "score": 1.0,
+ "content": "Chelsea Finn and Sergey Levine. Deep visual foresight for planning robot motion. In International",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 117,
+ 288,
+ 339,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 288,
+ 339,
+ 299
+ ],
+ "score": 1.0,
+ "content": "Conference on Robotics and Automation (ICRA), 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5,
+ "bbox_fs": [
+ 106,
+ 275,
+ 506,
+ 299
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 304,
+ 505,
+ 329
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 305,
+ 505,
+ 318
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 305,
+ 505,
+ 318
+ ],
+ "score": 1.0,
+ "content": "Chelsea Finn, Ian Goodfellow, and Sergey Levine. Unsupervised learning for physical interaction",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 116,
+ 316,
+ 450,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 316,
+ 450,
+ 329
+ ],
+ "score": 1.0,
+ "content": "through video prediction. In Neural Information Processing Systems (NIPS), 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15.5,
+ "bbox_fs": [
+ 106,
+ 305,
+ 505,
+ 329
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 334,
+ 504,
+ 369
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 506,
+ 348
+ ],
+ "score": 1.0,
+ "content": "Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 115,
+ 346,
+ 506,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 346,
+ 506,
+ 358
+ ],
+ "score": 1.0,
+ "content": "Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural informa-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 116,
+ 357,
+ 305,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 357,
+ 305,
+ 369
+ ],
+ "score": 1.0,
+ "content": "tion processing systems, pp. 2672–2680, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18,
+ "bbox_fs": [
+ 105,
+ 334,
+ 506,
+ 369
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 375,
+ 504,
+ 409
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 375,
+ 505,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 375,
+ 505,
+ 388
+ ],
+ "score": 1.0,
+ "content": "Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 115,
+ 386,
+ 506,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 386,
+ 506,
+ 399
+ ],
+ "score": 1.0,
+ "content": "mapping. In Computer vision and pattern recognition, 2006 IEEE computer society conference on,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 116,
+ 398,
+ 275,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 398,
+ 275,
+ 410
+ ],
+ "score": 1.0,
+ "content": "volume 2, pp. 1735–1742. IEEE, 2006.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21,
+ "bbox_fs": [
+ 105,
+ 375,
+ 506,
+ 410
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 415,
+ 378,
+ 428
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 415,
+ 379,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 415,
+ 379,
+ 430
+ ],
+ "score": 1.0,
+ "content": "Jakob Hohwy. The predictive mind. Oxford University Press, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 23,
+ "bbox_fs": [
+ 104,
+ 415,
+ 379,
+ 430
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 433,
+ 505,
+ 457
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 433,
+ 505,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 433,
+ 505,
+ 447
+ ],
+ "score": 1.0,
+ "content": "Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 116,
+ 445,
+ 307,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 445,
+ 307,
+ 456
+ ],
+ "score": 1.0,
+ "content": "conditional adversarial networks. CVPR, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5,
+ "bbox_fs": [
+ 105,
+ 433,
+ 505,
+ 456
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 463,
+ 503,
+ 487
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 464,
+ 505,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 464,
+ 505,
+ 477
+ ],
+ "score": 1.0,
+ "content": "Dinesh Jayaraman and Kristen Grauman. Learning image representations tied to ego-motion. In",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 115,
+ 474,
+ 494,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 474,
+ 494,
+ 487
+ ],
+ "score": 1.0,
+ "content": "Proceedings of the IEEE International Conference on Computer Vision, pp. 1413–1421, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 106,
+ 464,
+ 505,
+ 487
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 492,
+ 504,
+ 516
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 506,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 506,
+ 506
+ ],
+ "score": 1.0,
+ "content": "Nal Kalchbrenner, Aaron van den Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyals, Alex Graves,",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 117,
+ 504,
+ 473,
+ 517
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 504,
+ 473,
+ 517
+ ],
+ "score": 1.0,
+ "content": "and Koray Kavukcuoglu. Video pixel networks. arXiv preprint arXiv:1610.00527, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28.5,
+ "bbox_fs": [
+ 105,
+ 492,
+ 506,
+ 517
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 522,
+ 504,
+ 545
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 521,
+ 506,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 521,
+ 506,
+ 536
+ ],
+ "score": 1.0,
+ "content": "Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 115,
+ 533,
+ 214,
+ 544
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 533,
+ 214,
+ 544
+ ],
+ "score": 1.0,
+ "content": "arXiv:1312.6114, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.5,
+ "bbox_fs": [
+ 105,
+ 521,
+ 506,
+ 544
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 551,
+ 504,
+ 575
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "Kris M Kitani, Brian D Ziebart, James Andrew Bagnell, and Martial Hebert. Activity forecasting. In",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 116,
+ 562,
+ 409,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 562,
+ 409,
+ 576
+ ],
+ "score": 1.0,
+ "content": "European Conference on Computer Vision, pp. 201–214. Springer, 2012.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32.5,
+ "bbox_fs": [
+ 105,
+ 550,
+ 505,
+ 576
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 580,
+ 506,
+ 615
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 579,
+ 506,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 579,
+ 506,
+ 594
+ ],
+ "score": 1.0,
+ "content": "Dirk P Kroese, Reuven Y Rubinstein, Izack Cohen, Sergey Porotsky, and Thomas Taimre. Cross-",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 592,
+ 507,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 592,
+ 507,
+ 604
+ ],
+ "score": 1.0,
+ "content": "entropy method. In Encyclopedia of Operations Research and Management Science, pp. 326–333.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 115,
+ 602,
+ 181,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 602,
+ 181,
+ 617
+ ],
+ "score": 1.0,
+ "content": "Springer, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 105,
+ 579,
+ 507,
+ 617
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 621,
+ 505,
+ 645
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 621,
+ 506,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 621,
+ 506,
+ 634
+ ],
+ "score": 1.0,
+ "content": "Anders Boesen Lindbo Larsen, Søren Kaae Sønderby, Hugo Larochelle, and Ole Winther. Autoen-",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 116,
+ 632,
+ 393,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 632,
+ 393,
+ 645
+ ],
+ "score": 1.0,
+ "content": "coding beyond pixels using a learned similarity metric. ICML, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 37.5,
+ "bbox_fs": [
+ 106,
+ 621,
+ 506,
+ 645
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 650,
+ 503,
+ 674
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 504,
+ 664
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 504,
+ 664
+ ],
+ "score": 1.0,
+ "content": "Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 117,
+ 662,
+ 439,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 662,
+ 439,
+ 673
+ ],
+ "score": 1.0,
+ "content": "Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 39.5,
+ "bbox_fs": [
+ 106,
+ 649,
+ 504,
+ 673
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 679,
+ 503,
+ 703
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 679,
+ 505,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 679,
+ 505,
+ 693
+ ],
+ "score": 1.0,
+ "content": "Michael Mathieu, Camille Couprie, and Yann LeCun. Deep multi-scale video prediction beyond",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 116,
+ 692,
+ 356,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 692,
+ 356,
+ 702
+ ],
+ "score": 1.0,
+ "content": "mean square error. arXiv preprint arXiv:1511.05440, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 41.5,
+ "bbox_fs": [
+ 105,
+ 679,
+ 505,
+ 702
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 709,
+ 503,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 708,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 708,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Amy McGovern and Andrew G Barto. Automatic discovery of subgoals in reinforcement learning",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 115,
+ 721,
+ 271,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 721,
+ 271,
+ 732
+ ],
+ "score": 1.0,
+ "content": "using diverse density. In ICML, 2001.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43.5,
+ "bbox_fs": [
+ 106,
+ 708,
+ 505,
+ 732
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "Jan Hendrik Metzen. Online skill discovery using graph-based clustering. In European Workshop on",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 115,
+ 93,
+ 288,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 93,
+ 288,
+ 106
+ ],
+ "score": 1.0,
+ "content": "Reinforcement Learning, pp. 77–88, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 111,
+ 504,
+ 135
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 125
+ ],
+ "score": 1.0,
+ "content": "Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 115,
+ 123,
+ 214,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 123,
+ 214,
+ 134
+ ],
+ "score": 1.0,
+ "content": "arXiv:1411.1784, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 140,
+ 504,
+ 174
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 140,
+ 506,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 140,
+ 506,
+ 154
+ ],
+ "score": 1.0,
+ "content": "Hossein Mobahi, Ronan Collobert, and Jason Weston. Deep learning from temporal coherence in",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 115,
+ 150,
+ 506,
+ 166
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 150,
+ 506,
+ 166
+ ],
+ "score": 1.0,
+ "content": "video. In Proceedings of the 26th Annual International Conference on Machine Learning, pp.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 115,
+ 162,
+ 209,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 162,
+ 209,
+ 174
+ ],
+ "score": 1.0,
+ "content": "737–744. ACM, 2009.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 180,
+ 504,
+ 204
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 180,
+ 505,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 180,
+ 505,
+ 194
+ ],
+ "score": 1.0,
+ "content": "Alexander Neitz, Giambattista Parascandolo, Stefan Bauer, and Bernhard Schlkopf. Adaptive skip",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 115,
+ 192,
+ 425,
+ 204
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 192,
+ 425,
+ 204
+ ],
+ "score": 1.0,
+ "content": "intervals: Temporal abstraction for recurrent dynamical models. NIPS, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 210,
+ 505,
+ 244
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 210,
+ 506,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 210,
+ 506,
+ 222
+ ],
+ "score": 1.0,
+ "content": "Augustus Odena, Vincent Dumoulin, and Chris Olah. Deconvolution and checkerboard arti-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 115,
+ 221,
+ 506,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 221,
+ 506,
+ 234
+ ],
+ "score": 1.0,
+ "content": "facts. Distill, 2016. doi: 10.23915/distill.00003. URL http://distill.pub/2016/",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 117,
+ 234,
+ 235,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 234,
+ 235,
+ 243
+ ],
+ "score": 1.0,
+ "content": "deconv-checkerboard.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 250,
+ 504,
+ 284
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 250,
+ 505,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 250,
+ 505,
+ 262
+ ],
+ "score": 1.0,
+ "content": "Junhyuk Oh, Xiaoxiao Guo, Honglak Lee, Richard L Lewis, and Satinder Singh. Action-conditional",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 115,
+ 260,
+ 505,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 260,
+ 505,
+ 275
+ ],
+ "score": 1.0,
+ "content": "video prediction using deep networks in atari games. In Advances in Neural Information Processing",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 115,
+ 272,
+ 178,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 272,
+ 178,
+ 285
+ ],
+ "score": 1.0,
+ "content": "Systems, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 290,
+ 505,
+ 313
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 303
+ ],
+ "score": 1.0,
+ "content": "Aaron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 115,
+ 301,
+ 279,
+ 314
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 301,
+ 279,
+ 314
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:1601.06759, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 319,
+ 504,
+ 343
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 318,
+ 505,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 318,
+ 505,
+ 334
+ ],
+ "score": 1.0,
+ "content": "Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 115,
+ 331,
+ 468,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 331,
+ 468,
+ 343
+ ],
+ "score": 1.0,
+ "content": "convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 17.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 348,
+ 505,
+ 383
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 349,
+ 505,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 349,
+ 505,
+ 360
+ ],
+ "score": 1.0,
+ "content": "MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 116,
+ 360,
+ 505,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 360,
+ 505,
+ 373
+ ],
+ "score": 1.0,
+ "content": "Chopra. Video (language) modeling: a baseline for generative models of natural videos. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 115,
+ 372,
+ 249,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 372,
+ 249,
+ 383
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:1412.6604, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 389,
+ 505,
+ 412
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 388,
+ 506,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 388,
+ 506,
+ 403
+ ],
+ "score": 1.0,
+ "content": "Ozg ¨ ur¨ S¸ims¸ek and Andrew G Barto. Skill characterization based on betweenness. In Advances in",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 115,
+ 400,
+ 365,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 400,
+ 365,
+ 413
+ ],
+ "score": 1.0,
+ "content": "neural information processing systems, pp. 1497–1504, 2009.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 418,
+ 505,
+ 442
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 419,
+ 505,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 419,
+ 505,
+ 431
+ ],
+ "score": 1.0,
+ "content": "Richard S Sutton, Doina Precup, and Satinder Singh. Between mdps and semi-mdps: A framework",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 116,
+ 429,
+ 506,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 429,
+ 506,
+ 442
+ ],
+ "score": 1.0,
+ "content": "for temporal abstraction in reinforcement learning. Artificial intelligence, 112(1-2):181–211, 1999.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 447,
+ 505,
+ 481
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 447,
+ 506,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 447,
+ 506,
+ 461
+ ],
+ "score": 1.0,
+ "content": "Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based con-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 115,
+ 457,
+ 507,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 457,
+ 507,
+ 474
+ ],
+ "score": 1.0,
+ "content": "trol. In Intelligent Robots and Systems (IROS), 2012 IEEE/RSJ International Conference on, pp.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 116,
+ 470,
+ 218,
+ 481
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 470,
+ 218,
+ 481
+ ],
+ "score": 1.0,
+ "content": "5026–5033. IEEE, 2012.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 487,
+ 504,
+ 522
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 488,
+ 506,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 488,
+ 506,
+ 500
+ ],
+ "score": 1.0,
+ "content": "Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 115,
+ 499,
+ 507,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 499,
+ 507,
+ 512
+ ],
+ "score": 1.0,
+ "content": "image generation with pixelcnn decoders. In Advances in Neural Information Processing Systems,",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 115,
+ 511,
+ 207,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 511,
+ 207,
+ 522
+ ],
+ "score": 1.0,
+ "content": "pp. 4790–4798, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 528,
+ 505,
+ 551
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": "Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. In",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 539,
+ 415,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 539,
+ 415,
+ 552
+ ],
+ "score": 1.0,
+ "content": "Advances In Neural Information Processing Systems, pp. 613–621, 2016a.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 557,
+ 505,
+ 591
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 556,
+ 506,
+ 571
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 556,
+ 506,
+ 571
+ ],
+ "score": 1.0,
+ "content": "Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Anticipating visual representations from un-",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 568,
+ 506,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 568,
+ 506,
+ 581
+ ],
+ "score": 1.0,
+ "content": "labeled video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 115,
+ 579,
+ 196,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 579,
+ 196,
+ 591
+ ],
+ "score": 1.0,
+ "content": "pp. 98–106, 2016b.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 597,
+ 505,
+ 631
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 611
+ ],
+ "score": 1.0,
+ "content": "Jacob Walker, Carl Doersch, Abhinav Gupta, and Martial Hebert. An uncertain future: Forecasting",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 115,
+ 609,
+ 506,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 609,
+ 506,
+ 622
+ ],
+ "score": 1.0,
+ "content": "from static images using variational autoencoders. In European Conference on Computer Vision,",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 115,
+ 619,
+ 236,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 619,
+ 236,
+ 632
+ ],
+ "score": 1.0,
+ "content": "pp. 835–851. Springer, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 637,
+ 504,
+ 661
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "Xiaolong Wang, Ali Farhadi, and Abhinav Gupta. Actions˜ transformations. In Proceedings of the",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 116,
+ 649,
+ 460,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 649,
+ 460,
+ 662
+ ],
+ "score": 1.0,
+ "content": "IEEE conference on Computer Vision and Pattern Recognition, pp. 2658–2667, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 667,
+ 504,
+ 701
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 680
+ ],
+ "score": 1.0,
+ "content": "Tianfan Xue, Jiajun Wu, Katherine Bouman, and Bill Freeman. Visual dynamics: Probabilistic future",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 115,
+ 677,
+ 506,
+ 692
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 677,
+ 506,
+ 692
+ ],
+ "score": 1.0,
+ "content": "frame synthesis via cross convolutional networks. In Advances in Neural Information Processing",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 115,
+ 689,
+ 222,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 689,
+ 222,
+ 701
+ ],
+ "score": 1.0,
+ "content": "Systems, pp. 91–99, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 707,
+ 504,
+ 730
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 707,
+ 505,
+ 720
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 707,
+ 505,
+ 720
+ ],
+ "score": 1.0,
+ "content": "Tinghui Zhou, Shubham Tulsiani, Weilun Sun, Jitendra Malik, and Alexei A Efros. View synthesis",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 115,
+ 718,
+ 497,
+ 731
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 718,
+ 497,
+ 731
+ ],
+ "score": 1.0,
+ "content": "by appearance flow. In European conference on computer vision, pp. 286–301. Springer, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 45.5
+ }
+ ],
+ "page_idx": 10,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 310,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 765
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 765
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 15,
+ "width": 13
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "Jan Hendrik Metzen. Online skill discovery using graph-based clustering. In European Workshop on",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 115,
+ 93,
+ 288,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 93,
+ 288,
+ 106
+ ],
+ "score": 1.0,
+ "content": "Reinforcement Learning, pp. 77–88, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5,
+ "bbox_fs": [
+ 105,
+ 81,
+ 505,
+ 106
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 111,
+ 504,
+ 135
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 125
+ ],
+ "score": 1.0,
+ "content": "Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 115,
+ 123,
+ 214,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 123,
+ 214,
+ 134
+ ],
+ "score": 1.0,
+ "content": "arXiv:1411.1784, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.5,
+ "bbox_fs": [
+ 105,
+ 110,
+ 505,
+ 134
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 140,
+ 504,
+ 174
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 140,
+ 506,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 140,
+ 506,
+ 154
+ ],
+ "score": 1.0,
+ "content": "Hossein Mobahi, Ronan Collobert, and Jason Weston. Deep learning from temporal coherence in",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 115,
+ 150,
+ 506,
+ 166
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 150,
+ 506,
+ 166
+ ],
+ "score": 1.0,
+ "content": "video. In Proceedings of the 26th Annual International Conference on Machine Learning, pp.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 115,
+ 162,
+ 209,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 162,
+ 209,
+ 174
+ ],
+ "score": 1.0,
+ "content": "737–744. ACM, 2009.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5,
+ "bbox_fs": [
+ 105,
+ 140,
+ 506,
+ 174
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 180,
+ 504,
+ 204
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 180,
+ 505,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 180,
+ 505,
+ 194
+ ],
+ "score": 1.0,
+ "content": "Alexander Neitz, Giambattista Parascandolo, Stefan Bauer, and Bernhard Schlkopf. Adaptive skip",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 115,
+ 192,
+ 425,
+ 204
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 192,
+ 425,
+ 204
+ ],
+ "score": 1.0,
+ "content": "intervals: Temporal abstraction for recurrent dynamical models. NIPS, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5,
+ "bbox_fs": [
+ 106,
+ 180,
+ 505,
+ 204
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 210,
+ 505,
+ 244
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 210,
+ 506,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 210,
+ 506,
+ 222
+ ],
+ "score": 1.0,
+ "content": "Augustus Odena, Vincent Dumoulin, and Chris Olah. Deconvolution and checkerboard arti-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 115,
+ 221,
+ 506,
+ 234
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 221,
+ 506,
+ 234
+ ],
+ "score": 1.0,
+ "content": "facts. Distill, 2016. doi: 10.23915/distill.00003. URL http://distill.pub/2016/",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 117,
+ 234,
+ 235,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 234,
+ 235,
+ 243
+ ],
+ "score": 1.0,
+ "content": "deconv-checkerboard.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10,
+ "bbox_fs": [
+ 106,
+ 210,
+ 506,
+ 243
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 250,
+ 504,
+ 284
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 250,
+ 505,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 250,
+ 505,
+ 262
+ ],
+ "score": 1.0,
+ "content": "Junhyuk Oh, Xiaoxiao Guo, Honglak Lee, Richard L Lewis, and Satinder Singh. Action-conditional",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 115,
+ 260,
+ 505,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 260,
+ 505,
+ 275
+ ],
+ "score": 1.0,
+ "content": "video prediction using deep networks in atari games. In Advances in Neural Information Processing",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 115,
+ 272,
+ 178,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 272,
+ 178,
+ 285
+ ],
+ "score": 1.0,
+ "content": "Systems, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13,
+ "bbox_fs": [
+ 106,
+ 250,
+ 505,
+ 285
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 290,
+ 505,
+ 313
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 505,
+ 303
+ ],
+ "score": 1.0,
+ "content": "Aaron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 115,
+ 301,
+ 279,
+ 314
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 301,
+ 279,
+ 314
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint arXiv:1601.06759, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15.5,
+ "bbox_fs": [
+ 106,
+ 291,
+ 505,
+ 314
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 319,
+ 504,
+ 343
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 318,
+ 505,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 318,
+ 505,
+ 334
+ ],
+ "score": 1.0,
+ "content": "Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 115,
+ 331,
+ 468,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 331,
+ 468,
+ 343
+ ],
+ "score": 1.0,
+ "content": "convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 17.5,
+ "bbox_fs": [
+ 106,
+ 318,
+ 505,
+ 343
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 348,
+ 505,
+ 383
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 349,
+ 505,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 349,
+ 505,
+ 360
+ ],
+ "score": 1.0,
+ "content": "MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 116,
+ 360,
+ 505,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 360,
+ 505,
+ 373
+ ],
+ "score": 1.0,
+ "content": "Chopra. Video (language) modeling: a baseline for generative models of natural videos. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 115,
+ 372,
+ 249,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 372,
+ 249,
+ 383
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:1412.6604, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20,
+ "bbox_fs": [
+ 107,
+ 349,
+ 505,
+ 383
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 389,
+ 505,
+ 412
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 388,
+ 506,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 388,
+ 506,
+ 403
+ ],
+ "score": 1.0,
+ "content": "Ozg ¨ ur¨ S¸ims¸ek and Andrew G Barto. Skill characterization based on betweenness. In Advances in",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 115,
+ 400,
+ 365,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 400,
+ 365,
+ 413
+ ],
+ "score": 1.0,
+ "content": "neural information processing systems, pp. 1497–1504, 2009.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.5,
+ "bbox_fs": [
+ 106,
+ 388,
+ 506,
+ 413
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 418,
+ 505,
+ 442
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 419,
+ 505,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 419,
+ 505,
+ 431
+ ],
+ "score": 1.0,
+ "content": "Richard S Sutton, Doina Precup, and Satinder Singh. Between mdps and semi-mdps: A framework",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 116,
+ 429,
+ 506,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 429,
+ 506,
+ 442
+ ],
+ "score": 1.0,
+ "content": "for temporal abstraction in reinforcement learning. Artificial intelligence, 112(1-2):181–211, 1999.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5,
+ "bbox_fs": [
+ 106,
+ 419,
+ 506,
+ 442
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 447,
+ 505,
+ 481
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 447,
+ 506,
+ 461
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 447,
+ 506,
+ 461
+ ],
+ "score": 1.0,
+ "content": "Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based con-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 115,
+ 457,
+ 507,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 457,
+ 507,
+ 474
+ ],
+ "score": 1.0,
+ "content": "trol. In Intelligent Robots and Systems (IROS), 2012 IEEE/RSJ International Conference on, pp.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 116,
+ 470,
+ 218,
+ 481
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 470,
+ 218,
+ 481
+ ],
+ "score": 1.0,
+ "content": "5026–5033. IEEE, 2012.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27,
+ "bbox_fs": [
+ 106,
+ 447,
+ 507,
+ 481
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 487,
+ 504,
+ 522
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 488,
+ 506,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 488,
+ 506,
+ 500
+ ],
+ "score": 1.0,
+ "content": "Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 115,
+ 499,
+ 507,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 499,
+ 507,
+ 512
+ ],
+ "score": 1.0,
+ "content": "image generation with pixelcnn decoders. In Advances in Neural Information Processing Systems,",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 115,
+ 511,
+ 207,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 511,
+ 207,
+ 522
+ ],
+ "score": 1.0,
+ "content": "pp. 4790–4798, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 106,
+ 488,
+ 507,
+ 522
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 528,
+ 505,
+ 551
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": "Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. In",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 539,
+ 415,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 539,
+ 415,
+ 552
+ ],
+ "score": 1.0,
+ "content": "Advances In Neural Information Processing Systems, pp. 613–621, 2016a.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32.5,
+ "bbox_fs": [
+ 105,
+ 527,
+ 505,
+ 552
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 557,
+ 505,
+ 591
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 556,
+ 506,
+ 571
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 556,
+ 506,
+ 571
+ ],
+ "score": 1.0,
+ "content": "Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Anticipating visual representations from un-",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 568,
+ 506,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 568,
+ 506,
+ 581
+ ],
+ "score": 1.0,
+ "content": "labeled video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 115,
+ 579,
+ 196,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 579,
+ 196,
+ 591
+ ],
+ "score": 1.0,
+ "content": "pp. 98–106, 2016b.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 105,
+ 556,
+ 506,
+ 591
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 597,
+ 505,
+ 631
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 505,
+ 611
+ ],
+ "score": 1.0,
+ "content": "Jacob Walker, Carl Doersch, Abhinav Gupta, and Martial Hebert. An uncertain future: Forecasting",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 115,
+ 609,
+ 506,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 609,
+ 506,
+ 622
+ ],
+ "score": 1.0,
+ "content": "from static images using variational autoencoders. In European Conference on Computer Vision,",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 115,
+ 619,
+ 236,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 619,
+ 236,
+ 632
+ ],
+ "score": 1.0,
+ "content": "pp. 835–851. Springer, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38,
+ "bbox_fs": [
+ 105,
+ 596,
+ 506,
+ 632
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 637,
+ 504,
+ 661
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "Xiaolong Wang, Ali Farhadi, and Abhinav Gupta. Actions˜ transformations. In Proceedings of the",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 116,
+ 649,
+ 460,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 649,
+ 460,
+ 662
+ ],
+ "score": 1.0,
+ "content": "IEEE conference on Computer Vision and Pattern Recognition, pp. 2658–2667, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 106,
+ 637,
+ 505,
+ 662
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 667,
+ 504,
+ 701
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 680
+ ],
+ "score": 1.0,
+ "content": "Tianfan Xue, Jiajun Wu, Katherine Bouman, and Bill Freeman. Visual dynamics: Probabilistic future",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 115,
+ 677,
+ 506,
+ 692
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 677,
+ 506,
+ 692
+ ],
+ "score": 1.0,
+ "content": "frame synthesis via cross convolutional networks. In Advances in Neural Information Processing",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 115,
+ 689,
+ 222,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 689,
+ 222,
+ 701
+ ],
+ "score": 1.0,
+ "content": "Systems, pp. 91–99, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43,
+ "bbox_fs": [
+ 106,
+ 666,
+ 506,
+ 701
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 707,
+ 504,
+ 730
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 707,
+ 505,
+ 720
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 707,
+ 505,
+ 720
+ ],
+ "score": 1.0,
+ "content": "Tinghui Zhou, Shubham Tulsiani, Weilun Sun, Jitendra Malik, and Alexei A Efros. View synthesis",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 115,
+ 718,
+ 497,
+ 731
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 718,
+ 497,
+ 731
+ ],
+ "score": 1.0,
+ "content": "by appearance flow. In European conference on computer vision, pp. 286–301. Springer, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 45.5,
+ "bbox_fs": [
+ 106,
+ 707,
+ 505,
+ 731
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 104,
+ 80,
+ 505,
+ 225
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 104,
+ 80,
+ 505,
+ 225
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 81,
+ 505,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 81,
+ 505,
+ 224
+ ],
+ "score": 0.962,
+ "type": "image",
+ "image_path": "77d54b0e7d23e2a5f211a986f42660665ad7067860e4a9d856ffea03a249a4ba.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 104,
+ 80,
+ 505,
+ 128.33333333333334
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 104,
+ 128.33333333333334,
+ 505,
+ 176.66666666666669
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 104,
+ 176.66666666666669,
+ 505,
+ 225.00000000000003
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 105,
+ 233,
+ 505,
+ 254
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 232,
+ 505,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 232,
+ 380,
+ 245
+ ],
+ "score": 1.0,
+ "content": "Figure 11: Training time network schematic. At test time, only the predictor",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 380,
+ 234,
+ 388,
+ 243
+ ],
+ "score": 0.82,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 389,
+ 232,
+ 433,
+ 245
+ ],
+ "score": 1.0,
+ "content": "is used, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 433,
+ 233,
+ 482,
+ 244
+ ],
+ "score": 0.92,
+ "content": "z \\sim \\mathcal { N } ( 0 , \\mathcal { Z } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 482,
+ 232,
+ 505,
+ 245
+ ],
+ "score": 1.0,
+ "content": ". Loss",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 243,
+ 234,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 243,
+ 234,
+ 254
+ ],
+ "score": 1.0,
+ "content": "terms (as used in Eq 10) are in red.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 263,
+ 506,
+ 297
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 262,
+ 506,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 262,
+ 506,
+ 276
+ ],
+ "score": 1.0,
+ "content": "In these appendices, we provide details omitted in the main text for space. Note that more supple-",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 275,
+ 505,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 275,
+ 505,
+ 286
+ ],
+ "score": 1.0,
+ "content": "mentary material, such as video examples, is hosted at: https://sites.google.com/view/",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 153,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 153,
+ 297
+ ],
+ "score": 1.0,
+ "content": "ta-pred",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 105,
+ 311,
+ 457,
+ 325
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 311,
+ 458,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 311,
+ 458,
+ 326
+ ],
+ "score": 1.0,
+ "content": "A LABEL SMOOTHING FOR TIME-AGNOSTIC CONDITIONAL GANS",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 334,
+ 505,
+ 390
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 505,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 117,
+ 348
+ ],
+ "score": 1.0,
+ "content": "In",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 117,
+ 335,
+ 137,
+ 347
+ ],
+ "score": 0.3,
+ "content": "\\operatorname { E q } 7",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 137,
+ 334,
+ 346,
+ 348
+ ],
+ "score": 1.0,
+ "content": ", we defined a time-agnostic CGAN loss that trained",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 347,
+ 335,
+ 356,
+ 345
+ ],
+ "score": 0.39,
+ "content": "\\mathrm { T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 356,
+ 334,
+ 505,
+ 348
+ ],
+ "score": 1.0,
+ "content": "discriminators, one corresponding to",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 344,
+ 505,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 344,
+ 173,
+ 358
+ ],
+ "score": 1.0,
+ "content": "each target time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 174,
+ 346,
+ 198,
+ 356
+ ],
+ "score": 0.89,
+ "content": "t \\in \\mathrm { T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 199,
+ 344,
+ 341,
+ 358
+ ],
+ "score": 1.0,
+ "content": ". In Eq 7, the only positives for the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 341,
+ 347,
+ 346,
+ 356
+ ],
+ "score": 0.78,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 347,
+ 344,
+ 415,
+ 358
+ ],
+ "score": 1.0,
+ "content": "-th discriminator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 416,
+ 346,
+ 428,
+ 357
+ ],
+ "score": 0.89,
+ "content": "D _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 429,
+ 344,
+ 505,
+ 358
+ ],
+ "score": 1.0,
+ "content": "were ground truth",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 357,
+ 505,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 357,
+ 257,
+ 369
+ ],
+ "score": 1.0,
+ "content": "frames that occured precisely at time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 257,
+ 358,
+ 262,
+ 367
+ ],
+ "score": 0.8,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 357,
+ 344,
+ 369
+ ],
+ "score": 1.0,
+ "content": "relative to the input",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 359,
+ 350,
+ 367
+ ],
+ "score": 0.72,
+ "content": "c",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 350,
+ 357,
+ 505,
+ 369
+ ],
+ "score": 1.0,
+ "content": ", i.e., discriminators are trained so that",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 107,
+ 366,
+ 506,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 367,
+ 145,
+ 380
+ ],
+ "score": 0.93,
+ "content": "D _ { t } ( c , x _ { t } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 145,
+ 366,
+ 213,
+ 381
+ ],
+ "score": 1.0,
+ "content": "would be 1, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 213,
+ 368,
+ 263,
+ 380
+ ],
+ "score": 0.92,
+ "content": "D _ { t } ( c , x _ { t ^ { \\prime } \\neq t } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 366,
+ 281,
+ 381
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 282,
+ 367,
+ 331,
+ 380
+ ],
+ "score": 0.93,
+ "content": "D _ { t } ( c , G ( c ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 331,
+ 366,
+ 506,
+ 381
+ ],
+ "score": 1.0,
+ "content": "would both be 0. However, in practice, we",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 378,
+ 267,
+ 391
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 378,
+ 267,
+ 391
+ ],
+ "score": 1.0,
+ "content": "use the following slightly different loss:",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 140,
+ 391,
+ 470,
+ 458
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 140,
+ 391,
+ 470,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 140,
+ 391,
+ 470,
+ 458
+ ],
+ "score": 0.93,
+ "content": "\\begin{array} { r l } & { G ^ { * } = \\underset { G } { \\arg \\operatorname* { m i n } } \\underset { t \\in \\mathcal { T } } { \\min } \\underset { D _ { t } } { \\operatorname* { m a x } } \\mathcal { L } _ { \\mathrm { c g a n } } ^ { t } ( G , D _ { t } ) , } \\\\ & { \\mathcal { L } _ { \\mathrm { c g a n } } ^ { t } ( G , D _ { t } ) = \\log D _ { t } ( c , x _ { t } ) + \\log ( 1 - D _ { t } ( c , G ( c ) ) ) + } \\\\ & { \\qquad \\quad \\sum _ { t ^ { \\prime } \\not = t } [ l _ { t , t ^ { \\prime } } \\log D _ { t } ( c , x _ { t ^ { \\prime } } ) + ( 1 - l _ { t , t ^ { \\prime } } ) \\log ( 1 - D _ { t } ( c , x _ { t ^ { \\prime } } ) ) ] , } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "4525ffd75e604ae849a8afb8894be0951568926697e2a3b057606d386734f216.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 15,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 140,
+ 391,
+ 470,
+ 413.3333333333333
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 140,
+ 413.3333333333333,
+ 470,
+ 435.66666666666663
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 140,
+ 435.66666666666663,
+ 470,
+ 457.99999999999994
+ ],
+ "spans": [],
+ "index": 16
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 517
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 505,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 161,
+ 473
+ ],
+ "score": 1.0,
+ "content": "where we set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 162,
+ 461,
+ 277,
+ 473
+ ],
+ "score": 0.92,
+ "content": "l _ { t , t ^ { \\prime } } = \\operatorname* { m a x } ( 0 , 1 - \\alpha | t - t ^ { \\prime } | )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 277,
+ 461,
+ 299,
+ 473
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 299,
+ 461,
+ 338,
+ 472
+ ],
+ "score": 0.88,
+ "content": "\\alpha = 0 . 2 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 338,
+ 461,
+ 484,
+ 473
+ ],
+ "score": 1.0,
+ "content": ". The first two terms are the same as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 484,
+ 462,
+ 505,
+ 473
+ ],
+ "score": 0.34,
+ "content": "\\mathrm { E q } ~ 7",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 472,
+ 506,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 472,
+ 146,
+ 485
+ ],
+ "score": 1.0,
+ "content": "— for the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 473,
+ 151,
+ 482
+ ],
+ "score": 0.75,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 151,
+ 472,
+ 236,
+ 485
+ ],
+ "score": 1.0,
+ "content": "-th discriminator, the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 237,
+ 474,
+ 242,
+ 482
+ ],
+ "score": 0.76,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 242,
+ 472,
+ 506,
+ 485
+ ],
+ "score": 1.0,
+ "content": "-th frame provides a positive, and the generated frame provides a",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 483,
+ 506,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 483,
+ 347,
+ 496
+ ],
+ "score": 1.0,
+ "content": "negative instance. The third term defines the loss for frames",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 347,
+ 485,
+ 369,
+ 496
+ ],
+ "score": 0.88,
+ "content": "x _ { t ^ { \\prime } \\neq t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 369,
+ 483,
+ 456,
+ 496
+ ],
+ "score": 1.0,
+ "content": ": frames close to time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 457,
+ 485,
+ 461,
+ 493
+ ],
+ "score": 0.78,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 461,
+ 483,
+ 506,
+ 496
+ ],
+ "score": 1.0,
+ "content": "are partial",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 492,
+ 506,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 146,
+ 509
+ ],
+ "score": 1.0,
+ "content": "positives",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 495,
+ 201,
+ 506
+ ],
+ "score": 0.89,
+ "content": "( 0 < l _ { t , t ^ { \\prime } } < 1 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 492,
+ 304,
+ 509
+ ],
+ "score": 1.0,
+ "content": ", and others are negatives",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 495,
+ 342,
+ 506
+ ],
+ "score": 0.86,
+ "content": "( l _ { t , t ^ { \\prime } } = 0 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 342,
+ 492,
+ 506,
+ 509
+ ],
+ "score": 1.0,
+ "content": ". This label smoothing stabilizes training",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 505,
+ 332,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 505,
+ 316,
+ 518
+ ],
+ "score": 1.0,
+ "content": "and ensures enough positives for each discriminator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 316,
+ 506,
+ 328,
+ 516
+ ],
+ "score": 0.88,
+ "content": "D _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 329,
+ 505,
+ 332,
+ 518
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 532,
+ 325,
+ 545
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 531,
+ 326,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 531,
+ 326,
+ 547
+ ],
+ "score": 1.0,
+ "content": "B TIME-AGNOSTIC CONDITIONAL VAES",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 556,
+ 505,
+ 601
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 557,
+ 506,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 557,
+ 376,
+ 569
+ ],
+ "score": 1.0,
+ "content": "Here, we discuss one alternative to the TAP CVAE formulation of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 376,
+ 557,
+ 396,
+ 568
+ ],
+ "score": 0.42,
+ "content": "\\operatorname { E q } 9",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 397,
+ 557,
+ 506,
+ 569
+ ],
+ "score": 1.0,
+ "content": "in Sec 3.4. Rather than re-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 568,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 568,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": "stricting the minimum-over-time to be over the log-likelihood term alone, why not take the minimum-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 579,
+ 505,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 579,
+ 505,
+ 591
+ ],
+ "score": 1.0,
+ "content": "over-time over the whole loss? In other words, the inference network would still have to only see one",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 590,
+ 368,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 590,
+ 368,
+ 601
+ ],
+ "score": 1.0,
+ "content": "frame as input, and this version of the TAP CVAE loss would be:",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 24.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 158,
+ 603,
+ 452,
+ 622
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 158,
+ 603,
+ 452,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 158,
+ 603,
+ 452,
+ 622
+ ],
+ "score": 0.91,
+ "content": "\\mathcal { L } _ { \\mathrm { c v a e } } ( G , \\phi ) = \\operatorname* { m i n } _ { t \\in \\mathbb { T } } \\left[ D _ { K L } ( q _ { \\phi } ( z | x _ { t } ) , p ( z ) ) - \\mathbb { E } _ { z \\sim q _ { \\phi } ( z | x _ { t } ) } \\ln p _ { G } ( x _ { t } | c , z ) \\right] .",
+ "type": "interline_equation",
+ "image_path": "96ba7cbc5c034d740ad55663a0ae632523b6d30631ac44d8d0bbbef5eaf52d26.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 27,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 158,
+ 603,
+ 452,
+ 622
+ ],
+ "spans": [],
+ "index": 27
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 625,
+ 505,
+ 669
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 624,
+ 506,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 624,
+ 506,
+ 637
+ ],
+ "score": 1.0,
+ "content": "Unfortunately, this would be computationally very expensive to train. Both the inference network",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 107,
+ 636,
+ 504,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 638,
+ 117,
+ 649
+ ],
+ "score": 0.82,
+ "content": "q _ { \\phi }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 118,
+ 636,
+ 192,
+ 649
+ ],
+ "score": 1.0,
+ "content": "and the generator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 637,
+ 201,
+ 646
+ ],
+ "score": 0.82,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 636,
+ 295,
+ 649
+ ],
+ "score": 1.0,
+ "content": "would have to process",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 295,
+ 636,
+ 309,
+ 648
+ ],
+ "score": 0.79,
+ "content": "| \\mathrm { T } |",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 309,
+ 636,
+ 483,
+ 649
+ ],
+ "score": 1.0,
+ "content": "frames. In comparison, the formulation of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 484,
+ 636,
+ 504,
+ 648
+ ],
+ "score": 0.41,
+ "content": "\\operatorname { E q } 9",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 646,
+ 506,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 646,
+ 135,
+ 660
+ ],
+ "score": 1.0,
+ "content": "allows",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 135,
+ 649,
+ 146,
+ 659
+ ],
+ "score": 0.86,
+ "content": "q _ { \\phi }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 146,
+ 646,
+ 164,
+ 660
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 164,
+ 647,
+ 173,
+ 657
+ ],
+ "score": 0.82,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 174,
+ 646,
+ 447,
+ 660
+ ],
+ "score": 1.0,
+ "content": "to process just one frame, and only error computation must be done",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 447,
+ 648,
+ 456,
+ 657
+ ],
+ "score": 0.48,
+ "content": "\\mathrm { T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 456,
+ 646,
+ 506,
+ 660
+ ],
+ "score": 1.0,
+ "content": "times, once",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 658,
+ 251,
+ 670
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 658,
+ 197,
+ 670
+ ],
+ "score": 1.0,
+ "content": "with each target frame",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 198,
+ 659,
+ 208,
+ 669
+ ],
+ "score": 0.84,
+ "content": "x _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 209,
+ 658,
+ 223,
+ 670
+ ],
+ "score": 1.0,
+ "content": "for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 224,
+ 658,
+ 248,
+ 668
+ ],
+ "score": 0.89,
+ "content": "t \\in \\mathrm { T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 248,
+ 658,
+ 251,
+ 670
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 685,
+ 336,
+ 698
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 684,
+ 336,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 684,
+ 336,
+ 700
+ ],
+ "score": 1.0,
+ "content": "C ARCHITECTURE AND TRAINING DETAILS",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 709,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 353,
+ 722
+ ],
+ "score": 1.0,
+ "content": "Predictor architecture. As indicated in Fig 11, the predictor",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 353,
+ 712,
+ 362,
+ 720
+ ],
+ "score": 0.86,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 362,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "has an encoder-decoder architecture.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 720,
+ 504,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 720,
+ 497,
+ 732
+ ],
+ "score": 1.0,
+ "content": "When our approach is used together with the conditional VAE (Sec 3.4), the conditional latent",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 497,
+ 722,
+ 504,
+ 730
+ ],
+ "score": 0.71,
+ "content": "z",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 33.5
+ }
+ ],
+ "page_idx": 11,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "score": 1.0,
+ "content": "12",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 104,
+ 80,
+ 505,
+ 225
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 104,
+ 80,
+ 505,
+ 225
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 81,
+ 505,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 81,
+ 505,
+ 224
+ ],
+ "score": 0.962,
+ "type": "image",
+ "image_path": "77d54b0e7d23e2a5f211a986f42660665ad7067860e4a9d856ffea03a249a4ba.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 104,
+ 80,
+ 505,
+ 128.33333333333334
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 104,
+ 128.33333333333334,
+ 505,
+ 176.66666666666669
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 104,
+ 176.66666666666669,
+ 505,
+ 225.00000000000003
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 105,
+ 233,
+ 505,
+ 254
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 232,
+ 505,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 232,
+ 380,
+ 245
+ ],
+ "score": 1.0,
+ "content": "Figure 11: Training time network schematic. At test time, only the predictor",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 380,
+ 234,
+ 388,
+ 243
+ ],
+ "score": 0.82,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 389,
+ 232,
+ 433,
+ 245
+ ],
+ "score": 1.0,
+ "content": "is used, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 433,
+ 233,
+ 482,
+ 244
+ ],
+ "score": 0.92,
+ "content": "z \\sim \\mathcal { N } ( 0 , \\mathcal { Z } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 482,
+ 232,
+ 505,
+ 245
+ ],
+ "score": 1.0,
+ "content": ". Loss",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 243,
+ 234,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 243,
+ 234,
+ 254
+ ],
+ "score": 1.0,
+ "content": "terms (as used in Eq 10) are in red.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 263,
+ 506,
+ 297
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 262,
+ 506,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 262,
+ 506,
+ 276
+ ],
+ "score": 1.0,
+ "content": "In these appendices, we provide details omitted in the main text for space. Note that more supple-",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 275,
+ 505,
+ 286
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 275,
+ 505,
+ 286
+ ],
+ "score": 1.0,
+ "content": "mentary material, such as video examples, is hosted at: https://sites.google.com/view/",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 153,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 153,
+ 297
+ ],
+ "score": 1.0,
+ "content": "ta-pred",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 105,
+ 262,
+ 506,
+ 297
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 105,
+ 311,
+ 457,
+ 325
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 311,
+ 458,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 311,
+ 458,
+ 326
+ ],
+ "score": 1.0,
+ "content": "A LABEL SMOOTHING FOR TIME-AGNOSTIC CONDITIONAL GANS",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 334,
+ 505,
+ 390
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 505,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 117,
+ 348
+ ],
+ "score": 1.0,
+ "content": "In",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 117,
+ 335,
+ 137,
+ 347
+ ],
+ "score": 0.3,
+ "content": "\\operatorname { E q } 7",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 137,
+ 334,
+ 346,
+ 348
+ ],
+ "score": 1.0,
+ "content": ", we defined a time-agnostic CGAN loss that trained",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 347,
+ 335,
+ 356,
+ 345
+ ],
+ "score": 0.39,
+ "content": "\\mathrm { T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 356,
+ 334,
+ 505,
+ 348
+ ],
+ "score": 1.0,
+ "content": "discriminators, one corresponding to",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 344,
+ 505,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 344,
+ 173,
+ 358
+ ],
+ "score": 1.0,
+ "content": "each target time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 174,
+ 346,
+ 198,
+ 356
+ ],
+ "score": 0.89,
+ "content": "t \\in \\mathrm { T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 199,
+ 344,
+ 341,
+ 358
+ ],
+ "score": 1.0,
+ "content": ". In Eq 7, the only positives for the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 341,
+ 347,
+ 346,
+ 356
+ ],
+ "score": 0.78,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 347,
+ 344,
+ 415,
+ 358
+ ],
+ "score": 1.0,
+ "content": "-th discriminator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 416,
+ 346,
+ 428,
+ 357
+ ],
+ "score": 0.89,
+ "content": "D _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 429,
+ 344,
+ 505,
+ 358
+ ],
+ "score": 1.0,
+ "content": "were ground truth",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 357,
+ 505,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 357,
+ 257,
+ 369
+ ],
+ "score": 1.0,
+ "content": "frames that occured precisely at time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 257,
+ 358,
+ 262,
+ 367
+ ],
+ "score": 0.8,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 357,
+ 344,
+ 369
+ ],
+ "score": 1.0,
+ "content": "relative to the input",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 359,
+ 350,
+ 367
+ ],
+ "score": 0.72,
+ "content": "c",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 350,
+ 357,
+ 505,
+ 369
+ ],
+ "score": 1.0,
+ "content": ", i.e., discriminators are trained so that",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 107,
+ 366,
+ 506,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 367,
+ 145,
+ 380
+ ],
+ "score": 0.93,
+ "content": "D _ { t } ( c , x _ { t } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 145,
+ 366,
+ 213,
+ 381
+ ],
+ "score": 1.0,
+ "content": "would be 1, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 213,
+ 368,
+ 263,
+ 380
+ ],
+ "score": 0.92,
+ "content": "D _ { t } ( c , x _ { t ^ { \\prime } \\neq t } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 366,
+ 281,
+ 381
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 282,
+ 367,
+ 331,
+ 380
+ ],
+ "score": 0.93,
+ "content": "D _ { t } ( c , G ( c ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 331,
+ 366,
+ 506,
+ 381
+ ],
+ "score": 1.0,
+ "content": "would both be 0. However, in practice, we",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 378,
+ 267,
+ 391
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 378,
+ 267,
+ 391
+ ],
+ "score": 1.0,
+ "content": "use the following slightly different loss:",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 11,
+ "bbox_fs": [
+ 105,
+ 334,
+ 506,
+ 391
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 140,
+ 391,
+ 470,
+ 458
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 140,
+ 391,
+ 470,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 140,
+ 391,
+ 470,
+ 458
+ ],
+ "score": 0.93,
+ "content": "\\begin{array} { r l } & { G ^ { * } = \\underset { G } { \\arg \\operatorname* { m i n } } \\underset { t \\in \\mathcal { T } } { \\min } \\underset { D _ { t } } { \\operatorname* { m a x } } \\mathcal { L } _ { \\mathrm { c g a n } } ^ { t } ( G , D _ { t } ) , } \\\\ & { \\mathcal { L } _ { \\mathrm { c g a n } } ^ { t } ( G , D _ { t } ) = \\log D _ { t } ( c , x _ { t } ) + \\log ( 1 - D _ { t } ( c , G ( c ) ) ) + } \\\\ & { \\qquad \\quad \\sum _ { t ^ { \\prime } \\not = t } [ l _ { t , t ^ { \\prime } } \\log D _ { t } ( c , x _ { t ^ { \\prime } } ) + ( 1 - l _ { t , t ^ { \\prime } } ) \\log ( 1 - D _ { t } ( c , x _ { t ^ { \\prime } } ) ) ] , } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "4525ffd75e604ae849a8afb8894be0951568926697e2a3b057606d386734f216.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 15,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 140,
+ 391,
+ 470,
+ 413.3333333333333
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 140,
+ 413.3333333333333,
+ 470,
+ 435.66666666666663
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 140,
+ 435.66666666666663,
+ 470,
+ 457.99999999999994
+ ],
+ "spans": [],
+ "index": 16
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 517
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 505,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 161,
+ 473
+ ],
+ "score": 1.0,
+ "content": "where we set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 162,
+ 461,
+ 277,
+ 473
+ ],
+ "score": 0.92,
+ "content": "l _ { t , t ^ { \\prime } } = \\operatorname* { m a x } ( 0 , 1 - \\alpha | t - t ^ { \\prime } | )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 277,
+ 461,
+ 299,
+ 473
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 299,
+ 461,
+ 338,
+ 472
+ ],
+ "score": 0.88,
+ "content": "\\alpha = 0 . 2 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 338,
+ 461,
+ 484,
+ 473
+ ],
+ "score": 1.0,
+ "content": ". The first two terms are the same as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 484,
+ 462,
+ 505,
+ 473
+ ],
+ "score": 0.34,
+ "content": "\\mathrm { E q } ~ 7",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 472,
+ 506,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 472,
+ 146,
+ 485
+ ],
+ "score": 1.0,
+ "content": "— for the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 473,
+ 151,
+ 482
+ ],
+ "score": 0.75,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 151,
+ 472,
+ 236,
+ 485
+ ],
+ "score": 1.0,
+ "content": "-th discriminator, the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 237,
+ 474,
+ 242,
+ 482
+ ],
+ "score": 0.76,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 242,
+ 472,
+ 506,
+ 485
+ ],
+ "score": 1.0,
+ "content": "-th frame provides a positive, and the generated frame provides a",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 483,
+ 506,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 483,
+ 347,
+ 496
+ ],
+ "score": 1.0,
+ "content": "negative instance. The third term defines the loss for frames",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 347,
+ 485,
+ 369,
+ 496
+ ],
+ "score": 0.88,
+ "content": "x _ { t ^ { \\prime } \\neq t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 369,
+ 483,
+ 456,
+ 496
+ ],
+ "score": 1.0,
+ "content": ": frames close to time",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 457,
+ 485,
+ 461,
+ 493
+ ],
+ "score": 0.78,
+ "content": "t",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 461,
+ 483,
+ 506,
+ 496
+ ],
+ "score": 1.0,
+ "content": "are partial",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 492,
+ 506,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 146,
+ 509
+ ],
+ "score": 1.0,
+ "content": "positives",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 495,
+ 201,
+ 506
+ ],
+ "score": 0.89,
+ "content": "( 0 < l _ { t , t ^ { \\prime } } < 1 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 492,
+ 304,
+ 509
+ ],
+ "score": 1.0,
+ "content": ", and others are negatives",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 495,
+ 342,
+ 506
+ ],
+ "score": 0.86,
+ "content": "( l _ { t , t ^ { \\prime } } = 0 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 342,
+ 492,
+ 506,
+ 509
+ ],
+ "score": 1.0,
+ "content": ". This label smoothing stabilizes training",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 505,
+ 332,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 505,
+ 316,
+ 518
+ ],
+ "score": 1.0,
+ "content": "and ensures enough positives for each discriminator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 316,
+ 506,
+ 328,
+ 516
+ ],
+ "score": 0.88,
+ "content": "D _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 329,
+ 505,
+ 332,
+ 518
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 19,
+ "bbox_fs": [
+ 105,
+ 461,
+ 506,
+ 518
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 532,
+ 325,
+ 545
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 531,
+ 326,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 531,
+ 326,
+ 547
+ ],
+ "score": 1.0,
+ "content": "B TIME-AGNOSTIC CONDITIONAL VAES",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 556,
+ 505,
+ 601
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 557,
+ 506,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 557,
+ 376,
+ 569
+ ],
+ "score": 1.0,
+ "content": "Here, we discuss one alternative to the TAP CVAE formulation of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 376,
+ 557,
+ 396,
+ 568
+ ],
+ "score": 0.42,
+ "content": "\\operatorname { E q } 9",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 397,
+ 557,
+ 506,
+ 569
+ ],
+ "score": 1.0,
+ "content": "in Sec 3.4. Rather than re-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 568,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 568,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": "stricting the minimum-over-time to be over the log-likelihood term alone, why not take the minimum-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 579,
+ 505,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 579,
+ 505,
+ 591
+ ],
+ "score": 1.0,
+ "content": "over-time over the whole loss? In other words, the inference network would still have to only see one",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 590,
+ 368,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 590,
+ 368,
+ 601
+ ],
+ "score": 1.0,
+ "content": "frame as input, and this version of the TAP CVAE loss would be:",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 24.5,
+ "bbox_fs": [
+ 105,
+ 557,
+ 506,
+ 601
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 158,
+ 603,
+ 452,
+ 622
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 158,
+ 603,
+ 452,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 158,
+ 603,
+ 452,
+ 622
+ ],
+ "score": 0.91,
+ "content": "\\mathcal { L } _ { \\mathrm { c v a e } } ( G , \\phi ) = \\operatorname* { m i n } _ { t \\in \\mathbb { T } } \\left[ D _ { K L } ( q _ { \\phi } ( z | x _ { t } ) , p ( z ) ) - \\mathbb { E } _ { z \\sim q _ { \\phi } ( z | x _ { t } ) } \\ln p _ { G } ( x _ { t } | c , z ) \\right] .",
+ "type": "interline_equation",
+ "image_path": "96ba7cbc5c034d740ad55663a0ae632523b6d30631ac44d8d0bbbef5eaf52d26.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 27,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 158,
+ 603,
+ 452,
+ 622
+ ],
+ "spans": [],
+ "index": 27
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 625,
+ 505,
+ 669
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 624,
+ 506,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 624,
+ 506,
+ 637
+ ],
+ "score": 1.0,
+ "content": "Unfortunately, this would be computationally very expensive to train. Both the inference network",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 107,
+ 636,
+ 504,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 638,
+ 117,
+ 649
+ ],
+ "score": 0.82,
+ "content": "q _ { \\phi }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 118,
+ 636,
+ 192,
+ 649
+ ],
+ "score": 1.0,
+ "content": "and the generator",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 637,
+ 201,
+ 646
+ ],
+ "score": 0.82,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 636,
+ 295,
+ 649
+ ],
+ "score": 1.0,
+ "content": "would have to process",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 295,
+ 636,
+ 309,
+ 648
+ ],
+ "score": 0.79,
+ "content": "| \\mathrm { T } |",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 309,
+ 636,
+ 483,
+ 649
+ ],
+ "score": 1.0,
+ "content": "frames. In comparison, the formulation of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 484,
+ 636,
+ 504,
+ 648
+ ],
+ "score": 0.41,
+ "content": "\\operatorname { E q } 9",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 646,
+ 506,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 646,
+ 135,
+ 660
+ ],
+ "score": 1.0,
+ "content": "allows",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 135,
+ 649,
+ 146,
+ 659
+ ],
+ "score": 0.86,
+ "content": "q _ { \\phi }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 146,
+ 646,
+ 164,
+ 660
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 164,
+ 647,
+ 173,
+ 657
+ ],
+ "score": 0.82,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 174,
+ 646,
+ 447,
+ 660
+ ],
+ "score": 1.0,
+ "content": "to process just one frame, and only error computation must be done",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 447,
+ 648,
+ 456,
+ 657
+ ],
+ "score": 0.48,
+ "content": "\\mathrm { T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 456,
+ 646,
+ 506,
+ 660
+ ],
+ "score": 1.0,
+ "content": "times, once",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 658,
+ 251,
+ 670
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 658,
+ 197,
+ 670
+ ],
+ "score": 1.0,
+ "content": "with each target frame",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 198,
+ 659,
+ 208,
+ 669
+ ],
+ "score": 0.84,
+ "content": "x _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 209,
+ 658,
+ 223,
+ 670
+ ],
+ "score": 1.0,
+ "content": "for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 224,
+ 658,
+ 248,
+ 668
+ ],
+ "score": 0.89,
+ "content": "t \\in \\mathrm { T }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 248,
+ 658,
+ 251,
+ 670
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 29.5,
+ "bbox_fs": [
+ 105,
+ 624,
+ 506,
+ 670
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 685,
+ 336,
+ 698
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 684,
+ 336,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 684,
+ 336,
+ 700
+ ],
+ "score": 1.0,
+ "content": "C ARCHITECTURE AND TRAINING DETAILS",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 709,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 353,
+ 722
+ ],
+ "score": 1.0,
+ "content": "Predictor architecture. As indicated in Fig 11, the predictor",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 353,
+ 712,
+ 362,
+ 720
+ ],
+ "score": 0.86,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 362,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "has an encoder-decoder architecture.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 720,
+ 504,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 720,
+ 497,
+ 732
+ ],
+ "score": 1.0,
+ "content": "When our approach is used together with the conditional VAE (Sec 3.4), the conditional latent",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 497,
+ 722,
+ 504,
+ 730
+ ],
+ "score": 0.71,
+ "content": "z",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 215,
+ 506,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 215,
+ 506,
+ 227
+ ],
+ "score": 1.0,
+ "content": "is appended at the bottleneck between the encoder and the decoder. The encoder produces a 128-",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 226,
+ 505,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 226,
+ 277,
+ 237
+ ],
+ "score": 1.0,
+ "content": "dimensional code, and the VAE latent code",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 277,
+ 228,
+ 284,
+ 236
+ ],
+ "score": 0.76,
+ "content": "z",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 285,
+ 226,
+ 505,
+ 237
+ ],
+ "score": 1.0,
+ "content": "is 32-dimensional, so the overall size of the input to the",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 237,
+ 259,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 237,
+ 259,
+ 248
+ ],
+ "score": 1.0,
+ "content": "decoder is 160 when the VAE is used.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 33.5,
+ "bbox_fs": [
+ 105,
+ 709,
+ 506,
+ 732
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 135,
+ 78,
+ 477,
+ 168
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 135,
+ 78,
+ 477,
+ 168
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 135,
+ 78,
+ 477,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 135,
+ 78,
+ 477,
+ 168
+ ],
+ "score": 0.968,
+ "type": "image",
+ "image_path": "f8856cd9f2f3798eee429753195214b113ada59861c91af370862789af7f7cd3.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 135,
+ 78,
+ 477,
+ 108.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 135,
+ 108.0,
+ 477,
+ 138.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 135,
+ 138.0,
+ 477,
+ 168.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 177,
+ 505,
+ 198
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 176,
+ 506,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 176,
+ 506,
+ 189
+ ],
+ "score": 1.0,
+ "content": "Figure 12: The predictor produces an appearance flowfield which is used to produce flow-warped input frames,",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 187,
+ 456,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 187,
+ 456,
+ 198
+ ],
+ "score": 1.0,
+ "content": "which together with a frame of new pixels are masked and averaged to produce the output frame.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 215,
+ 505,
+ 248
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 215,
+ 506,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 215,
+ 506,
+ 227
+ ],
+ "score": 1.0,
+ "content": "is appended at the bottleneck between the encoder and the decoder. The encoder produces a 128-",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 226,
+ 505,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 226,
+ 277,
+ 237
+ ],
+ "score": 1.0,
+ "content": "dimensional code, and the VAE latent code",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 277,
+ 228,
+ 284,
+ 236
+ ],
+ "score": 0.76,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 285,
+ 226,
+ 505,
+ 237
+ ],
+ "score": 1.0,
+ "content": "is 32-dimensional, so the overall size of the input to the",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 237,
+ 259,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 237,
+ 259,
+ 248
+ ],
+ "score": 1.0,
+ "content": "decoder is 160 when the VAE is used.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 254,
+ 505,
+ 320
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 267
+ ],
+ "score": 1.0,
+ "content": "The encoder inside the predictor uses a series of 4x4convolution-batchnorm-LeakyReLU(0.2) blocks",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 265,
+ 506,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 265,
+ 253,
+ 277
+ ],
+ "score": 1.0,
+ "content": "to reduce the 64x64 input image as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 253,
+ 265,
+ 483,
+ 276
+ ],
+ "score": 0.83,
+ "content": "3 \\mathrm { x 6 4 \\mathrm { x 6 4 } } { } 3 2 \\mathrm { x } 3 2 \\mathrm { x } 3 2 { } 6 4 \\mathrm { x } 1 6 \\mathrm { x } 1 6 { } 1 2 8 \\mathrm { x } 8 \\mathrm { x } 8 { } 2 5 6 \\mathrm { x } 4 \\mathrm { x } 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 484,
+ 265,
+ 506,
+ 277
+ ],
+ "score": 1.0,
+ "content": ". For",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 276,
+ 504,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 276,
+ 466,
+ 288
+ ],
+ "score": 1.0,
+ "content": "bidirectional prediction, this is repeated for both images and concatenated to produce a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 467,
+ 276,
+ 504,
+ 286
+ ],
+ "score": 0.71,
+ "content": "5 1 2 \\mathrm { x } 4 \\mathrm { x } 4",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 287,
+ 505,
+ 298
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 287,
+ 505,
+ 298
+ ],
+ "score": 1.0,
+ "content": "feature map. Finally, a small convolution-ReLU-convolution head reduces this to a 128-dimensional",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 298,
+ 506,
+ 310
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 298,
+ 506,
+ 310
+ ],
+ "score": 1.0,
+ "content": "vector. Except for this last head subnetwork, this architecture is identical to the DCGAN discrimina-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 309,
+ 210,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 210,
+ 320
+ ],
+ "score": 1.0,
+ "content": "tor (Radford et al., 2015).",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 10.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 325,
+ 505,
+ 403
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 325,
+ 506,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 325,
+ 506,
+ 338
+ ],
+ "score": 1.0,
+ "content": "The decoder inside the predictor uses a DCGAN-based architecture (Radford et al., 2015).",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 336,
+ 507,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 336,
+ 391,
+ 350
+ ],
+ "score": 1.0,
+ "content": "The input vector is reshaped by a transposed convolution into 256",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 392,
+ 337,
+ 412,
+ 347
+ ],
+ "score": 0.61,
+ "content": "4 \\times 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 412,
+ 336,
+ 507,
+ 350
+ ],
+ "score": 1.0,
+ "content": "maps. This is subse-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 347,
+ 506,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 347,
+ 506,
+ 361
+ ],
+ "score": 1.0,
+ "content": "quently processed a series of bilinear upsampling-5x5convolution-batchnorm-ReLU blocks as",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 358,
+ 505,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 358,
+ 285,
+ 369
+ ],
+ "score": 0.91,
+ "content": "2 5 6 \\times 4 \\times 4 - 1 2 8 \\times 8 \\times 8 6 4 \\mathrm { x } 3 2 \\mathrm { x } 3 2 3 \\mathrm { x } 6 4 \\mathrm { x } \\bar { 6 } 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 285,
+ 358,
+ 505,
+ 371
+ ],
+ "score": 1.0,
+ "content": ". For the last block, a Tanh activation is used in place",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 370,
+ 505,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 370,
+ 505,
+ 382
+ ],
+ "score": 1.0,
+ "content": "of ReLU to keep outputs in [-1,1]. Compared to (Radford et al., 2015), the main difference is that",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 381,
+ 506,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 381,
+ 506,
+ 393
+ ],
+ "score": 1.0,
+ "content": "transposed convolutions are replaced by upsampling-5x5 convolution blocks, which aids in faster",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 392,
+ 333,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 333,
+ 404
+ ],
+ "score": 1.0,
+ "content": "learning with fewer image artifacts (Odena et al., 2016).",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 408,
+ 505,
+ 475
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 505,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 159,
+ 421
+ ],
+ "score": 1.0,
+ "content": "As shown in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 159,
+ 409,
+ 187,
+ 420
+ ],
+ "score": 0.25,
+ "content": "\\mathrm { F i g } 1 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 187,
+ 408,
+ 505,
+ 421
+ ],
+ "score": 1.0,
+ "content": ", our predictor produces three sets of outputs: (a) one frame of new pixels, (b)",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 506,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 336,
+ 432
+ ],
+ "score": 1.0,
+ "content": "“appearance flow” (Zhou et al., 2016) maps that warp the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 336,
+ 420,
+ 350,
+ 431
+ ],
+ "score": 0.9,
+ "content": "| C |",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 351,
+ 419,
+ 444,
+ 432
+ ],
+ "score": 1.0,
+ "content": "input image(s), and (c)",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 445,
+ 420,
+ 477,
+ 431
+ ],
+ "score": 0.92,
+ "content": "| C | + 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 477,
+ 419,
+ 506,
+ 432
+ ],
+ "score": 1.0,
+ "content": "masks",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 430,
+ 506,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 430,
+ 506,
+ 443
+ ],
+ "score": 1.0,
+ "content": "(summing to 1 at each pixel) that combine the warped input images and the new pixels frame to",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 442,
+ 505,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 505,
+ 453
+ ],
+ "score": 1.0,
+ "content": "produce the final output. To produce these three outputs, we use three decoders that all have the",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 453,
+ 504,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 453,
+ 504,
+ 464
+ ],
+ "score": 1.0,
+ "content": "same architecture as above, except that the final output is shaped appropriately—the appearance flow",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 463,
+ 501,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 463,
+ 179,
+ 476
+ ],
+ "score": 1.0,
+ "content": "decoder produces",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 179,
+ 463,
+ 224,
+ 475
+ ],
+ "score": 0.87,
+ "content": "\\vert C \\vert \\mathrm { x 6 4 x 6 4 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 224,
+ 463,
+ 401,
+ 476
+ ],
+ "score": 1.0,
+ "content": "flowfields, and the masks decoder produces",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 401,
+ 463,
+ 470,
+ 475
+ ],
+ "score": 0.89,
+ "content": "( | C | + 1 ) \\mathrm { x } 6 4 \\mathrm { x } 6 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 470,
+ 463,
+ 501,
+ 476
+ ],
+ "score": 1.0,
+ "content": "masks.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 23.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 493,
+ 505,
+ 560
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 506,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 506,
+ 506
+ ],
+ "score": 1.0,
+ "content": "Discriminator and Recognition Network The VAE recognition network and the discriminator",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 504,
+ 505,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 504,
+ 505,
+ 518
+ ],
+ "score": 1.0,
+ "content": "both use similar architectures to the predictor encoder above. Only the input layer and the output layer",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 516,
+ 505,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 516,
+ 326,
+ 529
+ ],
+ "score": 1.0,
+ "content": "are changed as follows: (a) The discriminator accepts",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 326,
+ 516,
+ 357,
+ 528
+ ],
+ "score": 0.92,
+ "content": "| C | + 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 358,
+ 516,
+ 505,
+ 529
+ ],
+ "score": 1.0,
+ "content": "images as input and the recognition",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 504,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 173,
+ 540
+ ],
+ "score": 1.0,
+ "content": "network accepts",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 174,
+ 528,
+ 212,
+ 539
+ ],
+ "score": 0.91,
+ "content": "| C | + | T |",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 527,
+ 490,
+ 540
+ ],
+ "score": 1.0,
+ "content": "images (the full video) as input. (b) The discriminator head produces",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 490,
+ 527,
+ 504,
+ 539
+ ],
+ "score": 0.87,
+ "content": "| T |",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 551
+ ],
+ "score": 1.0,
+ "content": "logits (one corresponding to each target time), and the recognition network produces a 32-dimensional",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 550,
+ 181,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 550,
+ 181,
+ 561
+ ],
+ "score": 1.0,
+ "content": "conditional latent.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 579,
+ 505,
+ 624
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 580,
+ 505,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 580,
+ 505,
+ 591
+ ],
+ "score": 1.0,
+ "content": "Training. We found it beneficial to initialize the decoder by training it first as an unconditional",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 591,
+ 505,
+ 603
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 591,
+ 505,
+ 603
+ ],
+ "score": 1.0,
+ "content": "frame generator on frames from the training videos. For this pretraining, we use learning rate 0.0001",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 601,
+ 505,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 601,
+ 505,
+ 614
+ ],
+ "score": 1.0,
+ "content": "for 10 epochs with batch size 64 and Adam optimizer. Thereafter, for training, we use learning rate",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 613,
+ 359,
+ 625
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 613,
+ 359,
+ 625
+ ],
+ "score": 1.0,
+ "content": "0.0001 for 200 epochs with batch size 64 and Adam optimizer.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 34.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 647,
+ 228,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 646,
+ 230,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 646,
+ 230,
+ 662
+ ],
+ "score": 1.0,
+ "content": "D DATA GENERATION",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 677,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 109,
+ 677,
+ 506,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 109,
+ 677,
+ 506,
+ 689
+ ],
+ "score": 1.0,
+ "content": "To generate the data, we use a cross-entropy method (CEM)-based planner (Kroese et al., 2013;",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Ebert et al., 2017) in the MuJoCo (Todorov et al., 2012) environment with access to the physics",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "engine, which produces non-deterministic trajectories. The planner plans towards randomly provided",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 507,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 507,
+ 723
+ ],
+ "score": 1.0,
+ "content": "goals, but we use both successful and failed trajectories. Sample videos of episodes are shown at:",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 331,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 331,
+ 733
+ ],
+ "score": 1.0,
+ "content": "https://sites.google.com/view/ta-pred",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "page_idx": 12,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 310,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 14,
+ "width": 13
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 135,
+ 78,
+ 477,
+ 168
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 135,
+ 78,
+ 477,
+ 168
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 135,
+ 78,
+ 477,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 135,
+ 78,
+ 477,
+ 168
+ ],
+ "score": 0.968,
+ "type": "image",
+ "image_path": "f8856cd9f2f3798eee429753195214b113ada59861c91af370862789af7f7cd3.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 135,
+ 78,
+ 477,
+ 108.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 135,
+ 108.0,
+ 477,
+ 138.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 135,
+ 138.0,
+ 477,
+ 168.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 177,
+ 505,
+ 198
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 176,
+ 506,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 176,
+ 506,
+ 189
+ ],
+ "score": 1.0,
+ "content": "Figure 12: The predictor produces an appearance flowfield which is used to produce flow-warped input frames,",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 187,
+ 456,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 187,
+ 456,
+ 198
+ ],
+ "score": 1.0,
+ "content": "which together with a frame of new pixels are masked and averaged to produce the output frame.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ }
+ ],
+ "index": 2.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 215,
+ 505,
+ 248
+ ],
+ "lines": [],
+ "index": 6,
+ "bbox_fs": [
+ 105,
+ 215,
+ 506,
+ 248
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 254,
+ 505,
+ 320
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 267
+ ],
+ "score": 1.0,
+ "content": "The encoder inside the predictor uses a series of 4x4convolution-batchnorm-LeakyReLU(0.2) blocks",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 265,
+ 506,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 265,
+ 253,
+ 277
+ ],
+ "score": 1.0,
+ "content": "to reduce the 64x64 input image as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 253,
+ 265,
+ 483,
+ 276
+ ],
+ "score": 0.83,
+ "content": "3 \\mathrm { x 6 4 \\mathrm { x 6 4 } } { } 3 2 \\mathrm { x } 3 2 \\mathrm { x } 3 2 { } 6 4 \\mathrm { x } 1 6 \\mathrm { x } 1 6 { } 1 2 8 \\mathrm { x } 8 \\mathrm { x } 8 { } 2 5 6 \\mathrm { x } 4 \\mathrm { x } 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 484,
+ 265,
+ 506,
+ 277
+ ],
+ "score": 1.0,
+ "content": ". For",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 276,
+ 504,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 276,
+ 466,
+ 288
+ ],
+ "score": 1.0,
+ "content": "bidirectional prediction, this is repeated for both images and concatenated to produce a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 467,
+ 276,
+ 504,
+ 286
+ ],
+ "score": 0.71,
+ "content": "5 1 2 \\mathrm { x } 4 \\mathrm { x } 4",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 287,
+ 505,
+ 298
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 287,
+ 505,
+ 298
+ ],
+ "score": 1.0,
+ "content": "feature map. Finally, a small convolution-ReLU-convolution head reduces this to a 128-dimensional",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 298,
+ 506,
+ 310
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 298,
+ 506,
+ 310
+ ],
+ "score": 1.0,
+ "content": "vector. Except for this last head subnetwork, this architecture is identical to the DCGAN discrimina-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 309,
+ 210,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 210,
+ 320
+ ],
+ "score": 1.0,
+ "content": "tor (Radford et al., 2015).",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 10.5,
+ "bbox_fs": [
+ 105,
+ 254,
+ 506,
+ 320
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 325,
+ 505,
+ 403
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 325,
+ 506,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 325,
+ 506,
+ 338
+ ],
+ "score": 1.0,
+ "content": "The decoder inside the predictor uses a DCGAN-based architecture (Radford et al., 2015).",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 336,
+ 507,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 336,
+ 391,
+ 350
+ ],
+ "score": 1.0,
+ "content": "The input vector is reshaped by a transposed convolution into 256",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 392,
+ 337,
+ 412,
+ 347
+ ],
+ "score": 0.61,
+ "content": "4 \\times 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 412,
+ 336,
+ 507,
+ 350
+ ],
+ "score": 1.0,
+ "content": "maps. This is subse-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 347,
+ 506,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 347,
+ 506,
+ 361
+ ],
+ "score": 1.0,
+ "content": "quently processed a series of bilinear upsampling-5x5convolution-batchnorm-ReLU blocks as",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 358,
+ 505,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 358,
+ 285,
+ 369
+ ],
+ "score": 0.91,
+ "content": "2 5 6 \\times 4 \\times 4 - 1 2 8 \\times 8 \\times 8 6 4 \\mathrm { x } 3 2 \\mathrm { x } 3 2 3 \\mathrm { x } 6 4 \\mathrm { x } \\bar { 6 } 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 285,
+ 358,
+ 505,
+ 371
+ ],
+ "score": 1.0,
+ "content": ". For the last block, a Tanh activation is used in place",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 370,
+ 505,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 370,
+ 505,
+ 382
+ ],
+ "score": 1.0,
+ "content": "of ReLU to keep outputs in [-1,1]. Compared to (Radford et al., 2015), the main difference is that",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 381,
+ 506,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 381,
+ 506,
+ 393
+ ],
+ "score": 1.0,
+ "content": "transposed convolutions are replaced by upsampling-5x5 convolution blocks, which aids in faster",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 392,
+ 333,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 333,
+ 404
+ ],
+ "score": 1.0,
+ "content": "learning with fewer image artifacts (Odena et al., 2016).",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 17,
+ "bbox_fs": [
+ 105,
+ 325,
+ 507,
+ 404
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 408,
+ 505,
+ 475
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 505,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 159,
+ 421
+ ],
+ "score": 1.0,
+ "content": "As shown in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 159,
+ 409,
+ 187,
+ 420
+ ],
+ "score": 0.25,
+ "content": "\\mathrm { F i g } 1 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 187,
+ 408,
+ 505,
+ 421
+ ],
+ "score": 1.0,
+ "content": ", our predictor produces three sets of outputs: (a) one frame of new pixels, (b)",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 506,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 336,
+ 432
+ ],
+ "score": 1.0,
+ "content": "“appearance flow” (Zhou et al., 2016) maps that warp the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 336,
+ 420,
+ 350,
+ 431
+ ],
+ "score": 0.9,
+ "content": "| C |",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 351,
+ 419,
+ 444,
+ 432
+ ],
+ "score": 1.0,
+ "content": "input image(s), and (c)",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 445,
+ 420,
+ 477,
+ 431
+ ],
+ "score": 0.92,
+ "content": "| C | + 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 477,
+ 419,
+ 506,
+ 432
+ ],
+ "score": 1.0,
+ "content": "masks",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 430,
+ 506,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 430,
+ 506,
+ 443
+ ],
+ "score": 1.0,
+ "content": "(summing to 1 at each pixel) that combine the warped input images and the new pixels frame to",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 442,
+ 505,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 505,
+ 453
+ ],
+ "score": 1.0,
+ "content": "produce the final output. To produce these three outputs, we use three decoders that all have the",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 453,
+ 504,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 453,
+ 504,
+ 464
+ ],
+ "score": 1.0,
+ "content": "same architecture as above, except that the final output is shaped appropriately—the appearance flow",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 463,
+ 501,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 463,
+ 179,
+ 476
+ ],
+ "score": 1.0,
+ "content": "decoder produces",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 179,
+ 463,
+ 224,
+ 475
+ ],
+ "score": 0.87,
+ "content": "\\vert C \\vert \\mathrm { x 6 4 x 6 4 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 224,
+ 463,
+ 401,
+ 476
+ ],
+ "score": 1.0,
+ "content": "flowfields, and the masks decoder produces",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 401,
+ 463,
+ 470,
+ 475
+ ],
+ "score": 0.89,
+ "content": "( | C | + 1 ) \\mathrm { x } 6 4 \\mathrm { x } 6 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 470,
+ 463,
+ 501,
+ 476
+ ],
+ "score": 1.0,
+ "content": "masks.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 23.5,
+ "bbox_fs": [
+ 105,
+ 408,
+ 506,
+ 476
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 493,
+ 505,
+ 560
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 506,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 506,
+ 506
+ ],
+ "score": 1.0,
+ "content": "Discriminator and Recognition Network The VAE recognition network and the discriminator",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 504,
+ 505,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 504,
+ 505,
+ 518
+ ],
+ "score": 1.0,
+ "content": "both use similar architectures to the predictor encoder above. Only the input layer and the output layer",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 516,
+ 505,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 516,
+ 326,
+ 529
+ ],
+ "score": 1.0,
+ "content": "are changed as follows: (a) The discriminator accepts",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 326,
+ 516,
+ 357,
+ 528
+ ],
+ "score": 0.92,
+ "content": "| C | + 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 358,
+ 516,
+ 505,
+ 529
+ ],
+ "score": 1.0,
+ "content": "images as input and the recognition",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 504,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 173,
+ 540
+ ],
+ "score": 1.0,
+ "content": "network accepts",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 174,
+ 528,
+ 212,
+ 539
+ ],
+ "score": 0.91,
+ "content": "| C | + | T |",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 213,
+ 527,
+ 490,
+ 540
+ ],
+ "score": 1.0,
+ "content": "images (the full video) as input. (b) The discriminator head produces",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 490,
+ 527,
+ 504,
+ 539
+ ],
+ "score": 0.87,
+ "content": "| T |",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 505,
+ 551
+ ],
+ "score": 1.0,
+ "content": "logits (one corresponding to each target time), and the recognition network produces a 32-dimensional",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 550,
+ 181,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 550,
+ 181,
+ 561
+ ],
+ "score": 1.0,
+ "content": "conditional latent.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 29.5,
+ "bbox_fs": [
+ 105,
+ 493,
+ 506,
+ 561
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 579,
+ 505,
+ 624
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 580,
+ 505,
+ 591
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 580,
+ 505,
+ 591
+ ],
+ "score": 1.0,
+ "content": "Training. We found it beneficial to initialize the decoder by training it first as an unconditional",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 591,
+ 505,
+ 603
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 591,
+ 505,
+ 603
+ ],
+ "score": 1.0,
+ "content": "frame generator on frames from the training videos. For this pretraining, we use learning rate 0.0001",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 601,
+ 505,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 601,
+ 505,
+ 614
+ ],
+ "score": 1.0,
+ "content": "for 10 epochs with batch size 64 and Adam optimizer. Thereafter, for training, we use learning rate",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 613,
+ 359,
+ 625
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 613,
+ 359,
+ 625
+ ],
+ "score": 1.0,
+ "content": "0.0001 for 200 epochs with batch size 64 and Adam optimizer.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 34.5,
+ "bbox_fs": [
+ 105,
+ 580,
+ 505,
+ 625
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 647,
+ 228,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 646,
+ 230,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 646,
+ 230,
+ 662
+ ],
+ "score": 1.0,
+ "content": "D DATA GENERATION",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 677,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 109,
+ 677,
+ 506,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 109,
+ 677,
+ 506,
+ 689
+ ],
+ "score": 1.0,
+ "content": "To generate the data, we use a cross-entropy method (CEM)-based planner (Kroese et al., 2013;",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Ebert et al., 2017) in the MuJoCo (Todorov et al., 2012) environment with access to the physics",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "engine, which produces non-deterministic trajectories. The planner plans towards randomly provided",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 507,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 507,
+ 723
+ ],
+ "score": 1.0,
+ "content": "goals, but we use both successful and failed trajectories. Sample videos of episodes are shown at:",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 331,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 331,
+ 733
+ ],
+ "score": 1.0,
+ "content": "https://sites.google.com/view/ta-pred",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 40,
+ "bbox_fs": [
+ 105,
+ 677,
+ 507,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 105,
+ 81,
+ 483,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 79,
+ 485,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 79,
+ 485,
+ 96
+ ],
+ "score": 1.0,
+ "content": "E GENERALIZED MINIMUM WEIGHTS FOR INTERMEDIATE PREDICTION",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 107,
+ 505,
+ 163
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 506,
+ 121
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 322,
+ 121
+ ],
+ "score": 1.0,
+ "content": "In Sec 4, we briefly mentioned that the time preference",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 323,
+ 107,
+ 343,
+ 119
+ ],
+ "score": 0.91,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 105,
+ 506,
+ 121
+ ],
+ "score": 1.0,
+ "content": "for the generalized minimum loss during",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 117,
+ 507,
+ 131
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 117,
+ 507,
+ 131
+ ],
+ "score": 1.0,
+ "content": "intermediate frame prediction was set to 2/3 at the ends to 1 at the middle frame. In our experiments,",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 129,
+ 505,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 129,
+ 286,
+ 141
+ ],
+ "score": 1.0,
+ "content": "we set these weights heuristically as follows:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 287,
+ 129,
+ 306,
+ 141
+ ],
+ "score": 0.92,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 307,
+ 129,
+ 383,
+ 141
+ ],
+ "score": 1.0,
+ "content": "rises linearly from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 383,
+ 130,
+ 421,
+ 140
+ ],
+ "score": 0.89,
+ "content": "\\kappa = 0 . 6 6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 421,
+ 129,
+ 431,
+ 141
+ ],
+ "score": 1.0,
+ "content": "at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 432,
+ 130,
+ 455,
+ 140
+ ],
+ "score": 0.9,
+ "content": "t = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 455,
+ 129,
+ 505,
+ 141
+ ],
+ "score": 1.0,
+ "content": "to 1.0 at the",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 107,
+ 139,
+ 505,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 141,
+ 130,
+ 151
+ ],
+ "score": 0.88,
+ "content": "t = 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 130,
+ 139,
+ 212,
+ 154
+ ],
+ "score": 1.0,
+ "content": ", then stay at 1.0 for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 212,
+ 141,
+ 243,
+ 151
+ ],
+ "score": 0.86,
+ "content": "T - 1 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 243,
+ 139,
+ 374,
+ 154
+ ],
+ "score": 1.0,
+ "content": "frames. Then, starting from the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 374,
+ 141,
+ 406,
+ 152
+ ],
+ "score": 0.92,
+ "content": "( T - 5 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 407,
+ 139,
+ 505,
+ 154
+ ],
+ "score": 1.0,
+ "content": "-th frame, it would drop",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 151,
+ 506,
+ 164
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 151,
+ 149,
+ 164
+ ],
+ "score": 1.0,
+ "content": "linearly to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 149,
+ 153,
+ 156,
+ 161
+ ],
+ "score": 0.76,
+ "content": "\\kappa",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 156,
+ 151,
+ 210,
+ 164
+ ],
+ "score": 1.0,
+ "content": "once more at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 210,
+ 152,
+ 235,
+ 161
+ ],
+ "score": 0.9,
+ "content": "t = T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 235,
+ 151,
+ 396,
+ 164
+ ],
+ "score": 1.0,
+ "content": ". The only hyperparameter we tuned was",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 397,
+ 153,
+ 403,
+ 161
+ ],
+ "score": 0.76,
+ "content": "\\kappa",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 404,
+ 151,
+ 454,
+ 164
+ ],
+ "score": 1.0,
+ "content": "(search over",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 455,
+ 151,
+ 468,
+ 162
+ ],
+ "score": 0.31,
+ "content": "2 / 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 469,
+ 151,
+ 486,
+ 164
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 487,
+ 152,
+ 500,
+ 162
+ ],
+ "score": 0.52,
+ "content": "1 / 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 500,
+ 151,
+ 506,
+ 164
+ ],
+ "score": 1.0,
+ "content": ").",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 180,
+ 366,
+ 194
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 180,
+ 367,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 180,
+ 367,
+ 196
+ ],
+ "score": 1.0,
+ "content": "F BOTTLENECK DISCOVERY FREQUENCY SCORE",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 207,
+ 504,
+ 229
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 207,
+ 505,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 207,
+ 505,
+ 221
+ ],
+ "score": 1.0,
+ "content": "In Sec 4, we mentioned a bottleneck discovery frequency measure in the paragraph titled “Bottleneck",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 218,
+ 352,
+ 230
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 218,
+ 352,
+ 230
+ ],
+ "score": 1.0,
+ "content": "discovery frequency metric.” We now provide further details.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 235,
+ 505,
+ 301
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 235,
+ 506,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 235,
+ 506,
+ 248
+ ],
+ "score": 1.0,
+ "content": "Our proposed metric for two-object pushing quantifies how reliably the network is able to generate a",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 246,
+ 506,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 246,
+ 506,
+ 259
+ ],
+ "score": 1.0,
+ "content": "bottleneck state with one object being moved and the other being at its original position. The reason",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 257,
+ 505,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 257,
+ 505,
+ 270
+ ],
+ "score": 1.0,
+ "content": "that this state is of interest is that in two-object pushing, this may be reasonably assumed to be the",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 267,
+ 505,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 267,
+ 505,
+ 281
+ ],
+ "score": 1.0,
+ "content": "natural bottleneck, so we call this metric the bottleneck frequency. Even without this assumption",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 279,
+ 506,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 279,
+ 506,
+ 292
+ ],
+ "score": 1.0,
+ "content": "though, the metric quantifies the ability of our approach to generate predictions attached to this",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 290,
+ 315,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 290,
+ 315,
+ 303
+ ],
+ "score": 1.0,
+ "content": "consistent semantically coherent bidirectional state.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 11.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 307,
+ 505,
+ 438
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 307,
+ 506,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 307,
+ 506,
+ 320
+ ],
+ "score": 1.0,
+ "content": "To measure bottleneck frequency, we use a technique similar to the approach proposed in (Ebert",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 104,
+ 317,
+ 507,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 317,
+ 507,
+ 330
+ ],
+ "score": 1.0,
+ "content": "et al., 2017) for planning with a visual predictor. First, we train versions of “genmin” and “fix”",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 328,
+ 505,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 328,
+ 505,
+ 342
+ ],
+ "score": 1.0,
+ "content": "that synthesize predictions purely by appearance-flow-warping and masking inputs (as shown in the",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 104,
+ 338,
+ 506,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 338,
+ 506,
+ 354
+ ],
+ "score": 1.0,
+ "content": "scheme of Fig 12, but without pixel generation). Next, recall that we have access to the starting",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 351,
+ 505,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 351,
+ 505,
+ 363
+ ],
+ "score": 1.0,
+ "content": "and goal positions of objects since our dataset is synthetically generated. Thus, we can exploit this",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 362,
+ 505,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 362,
+ 505,
+ 374
+ ],
+ "score": 1.0,
+ "content": "and track where the pixels at ground truth object locations in start and goal images end up in the",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 373,
+ 505,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 373,
+ 505,
+ 385
+ ],
+ "score": 1.0,
+ "content": "prediction, i.e., where did each object move? This works as follows: we take the appearance flow",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 383,
+ 505,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 383,
+ 505,
+ 396
+ ],
+ "score": 1.0,
+ "content": "transformations and masks generated by the model internally (for application to input images to",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 395,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 395,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "generate prediction) and apply them instead to one-hot object location maps—these maps have value",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 406,
+ 505,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 406,
+ 505,
+ 418
+ ],
+ "score": 1.0,
+ "content": "one at the ground truth origin of the the objects and zero elsewhere. The output of this operation",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 505,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 505,
+ 430
+ ],
+ "score": 1.0,
+ "content": "is a probability map for each object indicating where it is located in the predictor’s bidirectional",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 179,
+ 440
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 179,
+ 440
+ ],
+ "score": 1.0,
+ "content": "prediction output.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 20.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 444,
+ 505,
+ 510
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 444,
+ 505,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 444,
+ 505,
+ 457
+ ],
+ "score": 1.0,
+ "content": "To compute the score, we then calculate the expected distance in pixels between the predicted",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 455,
+ 506,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 455,
+ 506,
+ 468
+ ],
+ "score": 1.0,
+ "content": "positions of each object and the bottleneck state. There are actually two possible candidates for this",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 465,
+ 506,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 506,
+ 479
+ ],
+ "score": 1.0,
+ "content": "bottleneck state: object 1 is moved first, or object 2 is moved first. We compute the expected distances",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 478,
+ 505,
+ 489
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 478,
+ 505,
+ 489
+ ],
+ "score": 1.0,
+ "content": "to both these bottleneck candidates and take the lower distance to be the score. This score does not",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 489,
+ 505,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 489,
+ 505,
+ 500
+ ],
+ "score": 1.0,
+ "content": "evaluate whether the semantically correct bottleneck was predicted (in cases where one object must",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 500,
+ 401,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 500,
+ 401,
+ 511
+ ],
+ "score": 1.0,
+ "content": "always be moved first to avoid collision, such as Fig 9 (left) example #1).",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 516,
+ 505,
+ 593
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 506,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 506,
+ 528
+ ],
+ "score": 1.0,
+ "content": "The lower this distance score, the higher the likelihood that the predicted output is actually a bottleneck",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 526,
+ 506,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 526,
+ 506,
+ 540
+ ],
+ "score": 1.0,
+ "content": "(“bottleneck frequency”). Fig 10 shows what happens when the threshold over the score is varied, for",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 104,
+ 537,
+ 507,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 537,
+ 507,
+ 550
+ ],
+ "score": 1.0,
+ "content": "our approach and the fixed-time baseline. Higher bottleneck frequency at lower threshold is desirable.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 549,
+ 505,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 549,
+ 326,
+ 562
+ ],
+ "score": 1.0,
+ "content": "As the figure shows, at a low threshold distance score",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 326,
+ 550,
+ 343,
+ 560
+ ],
+ "score": 0.79,
+ "content": "\\approx 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 549,
+ 484,
+ 562
+ ],
+ "score": 1.0,
+ "content": "pixels), our approach gets to about",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 485,
+ 549,
+ 505,
+ 560
+ ],
+ "score": 0.85,
+ "content": "58 \\%",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 560,
+ 505,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 357,
+ 572
+ ],
+ "score": 1.0,
+ "content": "bottleneck frequency while the fixed-time predictor gets about",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 357,
+ 560,
+ 372,
+ 570
+ ],
+ "score": 0.87,
+ "content": "0 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 372,
+ 560,
+ 505,
+ 572
+ ],
+ "score": 1.0,
+ "content": "frequency at this threshold. This",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 104,
+ 570,
+ 505,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 570,
+ 505,
+ 585
+ ],
+ "score": 1.0,
+ "content": "verifies that our approach produces predictions attached to semantically coherent low-uncertainty",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 581,
+ 181,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 581,
+ 181,
+ 594
+ ],
+ "score": 1.0,
+ "content": "bottleneck events.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 611,
+ 380,
+ 624
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 381,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 381,
+ 626
+ ],
+ "score": 1.0,
+ "content": "G HIERARCHICAL PLANNING EVALUTION METHOD",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 40
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 638,
+ 504,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "In Sec 4, we described a hierarchical planning approach using our predictions in the paragraph titled",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 648,
+ 429,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 429,
+ 661
+ ],
+ "score": 1.0,
+ "content": "“Hierarchical planning evaluation.” We describe this method in more detail here.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 41.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 666,
+ 505,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 506,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 506,
+ 679
+ ],
+ "score": 1.0,
+ "content": "We test the usefulness of the generated predictions as subgoals for planning multi-object pushing.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "Given a start and goal image, we produce an bidirectional prediction using our time-agnostic predictor",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 699
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 699
+ ],
+ "score": 1.0,
+ "content": "and feed it to a low-level planner that plans towards this prediction as its subgoal. For the low-level",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "planner we use the visual model-predictive control algorithm (“Visual MPC”) (Ebert et al., 2017)",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "which internally uses a fixed-time forward prediction model and sampling-based planning to solve",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 472,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 472,
+ 733
+ ],
+ "score": 1.0,
+ "content": "short-term and medium duration tasks. A more detailed description of this process follows.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 45.5
+ }
+ ],
+ "page_idx": 13,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 752,
+ 310,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 14,
+ "width": 13
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 105,
+ 81,
+ 483,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 79,
+ 485,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 79,
+ 485,
+ 96
+ ],
+ "score": 1.0,
+ "content": "E GENERALIZED MINIMUM WEIGHTS FOR INTERMEDIATE PREDICTION",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 107,
+ 505,
+ 163
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 506,
+ 121
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 322,
+ 121
+ ],
+ "score": 1.0,
+ "content": "In Sec 4, we briefly mentioned that the time preference",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 323,
+ 107,
+ 343,
+ 119
+ ],
+ "score": 0.91,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 105,
+ 506,
+ 121
+ ],
+ "score": 1.0,
+ "content": "for the generalized minimum loss during",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 117,
+ 507,
+ 131
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 117,
+ 507,
+ 131
+ ],
+ "score": 1.0,
+ "content": "intermediate frame prediction was set to 2/3 at the ends to 1 at the middle frame. In our experiments,",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 129,
+ 505,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 129,
+ 286,
+ 141
+ ],
+ "score": 1.0,
+ "content": "we set these weights heuristically as follows:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 287,
+ 129,
+ 306,
+ 141
+ ],
+ "score": 0.92,
+ "content": "w ( t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 307,
+ 129,
+ 383,
+ 141
+ ],
+ "score": 1.0,
+ "content": "rises linearly from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 383,
+ 130,
+ 421,
+ 140
+ ],
+ "score": 0.89,
+ "content": "\\kappa = 0 . 6 6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 421,
+ 129,
+ 431,
+ 141
+ ],
+ "score": 1.0,
+ "content": "at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 432,
+ 130,
+ 455,
+ 140
+ ],
+ "score": 0.9,
+ "content": "t = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 455,
+ 129,
+ 505,
+ 141
+ ],
+ "score": 1.0,
+ "content": "to 1.0 at the",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 107,
+ 139,
+ 505,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 141,
+ 130,
+ 151
+ ],
+ "score": 0.88,
+ "content": "t = 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 130,
+ 139,
+ 212,
+ 154
+ ],
+ "score": 1.0,
+ "content": ", then stay at 1.0 for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 212,
+ 141,
+ 243,
+ 151
+ ],
+ "score": 0.86,
+ "content": "T - 1 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 243,
+ 139,
+ 374,
+ 154
+ ],
+ "score": 1.0,
+ "content": "frames. Then, starting from the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 374,
+ 141,
+ 406,
+ 152
+ ],
+ "score": 0.92,
+ "content": "( T - 5 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 407,
+ 139,
+ 505,
+ 154
+ ],
+ "score": 1.0,
+ "content": "-th frame, it would drop",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 151,
+ 506,
+ 164
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 151,
+ 149,
+ 164
+ ],
+ "score": 1.0,
+ "content": "linearly to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 149,
+ 153,
+ 156,
+ 161
+ ],
+ "score": 0.76,
+ "content": "\\kappa",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 156,
+ 151,
+ 210,
+ 164
+ ],
+ "score": 1.0,
+ "content": "once more at",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 210,
+ 152,
+ 235,
+ 161
+ ],
+ "score": 0.9,
+ "content": "t = T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 235,
+ 151,
+ 396,
+ 164
+ ],
+ "score": 1.0,
+ "content": ". The only hyperparameter we tuned was",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 397,
+ 153,
+ 403,
+ 161
+ ],
+ "score": 0.76,
+ "content": "\\kappa",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 404,
+ 151,
+ 454,
+ 164
+ ],
+ "score": 1.0,
+ "content": "(search over",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 455,
+ 151,
+ 468,
+ 162
+ ],
+ "score": 0.31,
+ "content": "2 / 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 469,
+ 151,
+ 486,
+ 164
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 487,
+ 152,
+ 500,
+ 162
+ ],
+ "score": 0.52,
+ "content": "1 / 3",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 500,
+ 151,
+ 506,
+ 164
+ ],
+ "score": 1.0,
+ "content": ").",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3,
+ "bbox_fs": [
+ 105,
+ 105,
+ 507,
+ 164
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 180,
+ 366,
+ 194
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 180,
+ 367,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 180,
+ 367,
+ 196
+ ],
+ "score": 1.0,
+ "content": "F BOTTLENECK DISCOVERY FREQUENCY SCORE",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 207,
+ 504,
+ 229
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 207,
+ 505,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 207,
+ 505,
+ 221
+ ],
+ "score": 1.0,
+ "content": "In Sec 4, we mentioned a bottleneck discovery frequency measure in the paragraph titled “Bottleneck",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 218,
+ 352,
+ 230
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 218,
+ 352,
+ 230
+ ],
+ "score": 1.0,
+ "content": "discovery frequency metric.” We now provide further details.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5,
+ "bbox_fs": [
+ 106,
+ 207,
+ 505,
+ 230
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 235,
+ 505,
+ 301
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 235,
+ 506,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 235,
+ 506,
+ 248
+ ],
+ "score": 1.0,
+ "content": "Our proposed metric for two-object pushing quantifies how reliably the network is able to generate a",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 246,
+ 506,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 246,
+ 506,
+ 259
+ ],
+ "score": 1.0,
+ "content": "bottleneck state with one object being moved and the other being at its original position. The reason",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 257,
+ 505,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 257,
+ 505,
+ 270
+ ],
+ "score": 1.0,
+ "content": "that this state is of interest is that in two-object pushing, this may be reasonably assumed to be the",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 267,
+ 505,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 267,
+ 505,
+ 281
+ ],
+ "score": 1.0,
+ "content": "natural bottleneck, so we call this metric the bottleneck frequency. Even without this assumption",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 279,
+ 506,
+ 292
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 279,
+ 506,
+ 292
+ ],
+ "score": 1.0,
+ "content": "though, the metric quantifies the ability of our approach to generate predictions attached to this",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 290,
+ 315,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 290,
+ 315,
+ 303
+ ],
+ "score": 1.0,
+ "content": "consistent semantically coherent bidirectional state.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 11.5,
+ "bbox_fs": [
+ 105,
+ 235,
+ 506,
+ 303
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 307,
+ 505,
+ 438
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 307,
+ 506,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 307,
+ 506,
+ 320
+ ],
+ "score": 1.0,
+ "content": "To measure bottleneck frequency, we use a technique similar to the approach proposed in (Ebert",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 104,
+ 317,
+ 507,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 317,
+ 507,
+ 330
+ ],
+ "score": 1.0,
+ "content": "et al., 2017) for planning with a visual predictor. First, we train versions of “genmin” and “fix”",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 328,
+ 505,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 328,
+ 505,
+ 342
+ ],
+ "score": 1.0,
+ "content": "that synthesize predictions purely by appearance-flow-warping and masking inputs (as shown in the",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 104,
+ 338,
+ 506,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 338,
+ 506,
+ 354
+ ],
+ "score": 1.0,
+ "content": "scheme of Fig 12, but without pixel generation). Next, recall that we have access to the starting",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 351,
+ 505,
+ 363
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 351,
+ 505,
+ 363
+ ],
+ "score": 1.0,
+ "content": "and goal positions of objects since our dataset is synthetically generated. Thus, we can exploit this",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 362,
+ 505,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 362,
+ 505,
+ 374
+ ],
+ "score": 1.0,
+ "content": "and track where the pixels at ground truth object locations in start and goal images end up in the",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 373,
+ 505,
+ 385
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 373,
+ 505,
+ 385
+ ],
+ "score": 1.0,
+ "content": "prediction, i.e., where did each object move? This works as follows: we take the appearance flow",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 383,
+ 505,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 383,
+ 505,
+ 396
+ ],
+ "score": 1.0,
+ "content": "transformations and masks generated by the model internally (for application to input images to",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 395,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 395,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "generate prediction) and apply them instead to one-hot object location maps—these maps have value",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 406,
+ 505,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 406,
+ 505,
+ 418
+ ],
+ "score": 1.0,
+ "content": "one at the ground truth origin of the the objects and zero elsewhere. The output of this operation",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 505,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 505,
+ 430
+ ],
+ "score": 1.0,
+ "content": "is a probability map for each object indicating where it is located in the predictor’s bidirectional",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 179,
+ 440
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 179,
+ 440
+ ],
+ "score": 1.0,
+ "content": "prediction output.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 20.5,
+ "bbox_fs": [
+ 104,
+ 307,
+ 507,
+ 440
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 444,
+ 505,
+ 510
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 444,
+ 505,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 444,
+ 505,
+ 457
+ ],
+ "score": 1.0,
+ "content": "To compute the score, we then calculate the expected distance in pixels between the predicted",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 455,
+ 506,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 455,
+ 506,
+ 468
+ ],
+ "score": 1.0,
+ "content": "positions of each object and the bottleneck state. There are actually two possible candidates for this",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 465,
+ 506,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 506,
+ 479
+ ],
+ "score": 1.0,
+ "content": "bottleneck state: object 1 is moved first, or object 2 is moved first. We compute the expected distances",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 478,
+ 505,
+ 489
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 478,
+ 505,
+ 489
+ ],
+ "score": 1.0,
+ "content": "to both these bottleneck candidates and take the lower distance to be the score. This score does not",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 489,
+ 505,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 489,
+ 505,
+ 500
+ ],
+ "score": 1.0,
+ "content": "evaluate whether the semantically correct bottleneck was predicted (in cases where one object must",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 500,
+ 401,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 500,
+ 401,
+ 511
+ ],
+ "score": 1.0,
+ "content": "always be moved first to avoid collision, such as Fig 9 (left) example #1).",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 29.5,
+ "bbox_fs": [
+ 105,
+ 444,
+ 506,
+ 511
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 516,
+ 505,
+ 593
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 506,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 506,
+ 528
+ ],
+ "score": 1.0,
+ "content": "The lower this distance score, the higher the likelihood that the predicted output is actually a bottleneck",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 526,
+ 506,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 526,
+ 506,
+ 540
+ ],
+ "score": 1.0,
+ "content": "(“bottleneck frequency”). Fig 10 shows what happens when the threshold over the score is varied, for",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 104,
+ 537,
+ 507,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 537,
+ 507,
+ 550
+ ],
+ "score": 1.0,
+ "content": "our approach and the fixed-time baseline. Higher bottleneck frequency at lower threshold is desirable.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 549,
+ 505,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 549,
+ 326,
+ 562
+ ],
+ "score": 1.0,
+ "content": "As the figure shows, at a low threshold distance score",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 326,
+ 550,
+ 343,
+ 560
+ ],
+ "score": 0.79,
+ "content": "\\approx 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 549,
+ 484,
+ 562
+ ],
+ "score": 1.0,
+ "content": "pixels), our approach gets to about",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 485,
+ 549,
+ 505,
+ 560
+ ],
+ "score": 0.85,
+ "content": "58 \\%",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 560,
+ 505,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 357,
+ 572
+ ],
+ "score": 1.0,
+ "content": "bottleneck frequency while the fixed-time predictor gets about",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 357,
+ 560,
+ 372,
+ 570
+ ],
+ "score": 0.87,
+ "content": "0 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 372,
+ 560,
+ 505,
+ 572
+ ],
+ "score": 1.0,
+ "content": "frequency at this threshold. This",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 104,
+ 570,
+ 505,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 570,
+ 505,
+ 585
+ ],
+ "score": 1.0,
+ "content": "verifies that our approach produces predictions attached to semantically coherent low-uncertainty",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 581,
+ 181,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 581,
+ 181,
+ 594
+ ],
+ "score": 1.0,
+ "content": "bottleneck events.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 36,
+ "bbox_fs": [
+ 104,
+ 515,
+ 507,
+ 594
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 611,
+ 380,
+ 624
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 381,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 381,
+ 626
+ ],
+ "score": 1.0,
+ "content": "G HIERARCHICAL PLANNING EVALUTION METHOD",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 40
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 638,
+ 504,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "In Sec 4, we described a hierarchical planning approach using our predictions in the paragraph titled",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 648,
+ 429,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 429,
+ 661
+ ],
+ "score": 1.0,
+ "content": "“Hierarchical planning evaluation.” We describe this method in more detail here.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 41.5,
+ "bbox_fs": [
+ 105,
+ 637,
+ 505,
+ 661
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 666,
+ 505,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 506,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 506,
+ 679
+ ],
+ "score": 1.0,
+ "content": "We test the usefulness of the generated predictions as subgoals for planning multi-object pushing.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "Given a start and goal image, we produce an bidirectional prediction using our time-agnostic predictor",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 699
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 699
+ ],
+ "score": 1.0,
+ "content": "and feed it to a low-level planner that plans towards this prediction as its subgoal. For the low-level",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "planner we use the visual model-predictive control algorithm (“Visual MPC”) (Ebert et al., 2017)",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "which internally uses a fixed-time forward prediction model and sampling-based planning to solve",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 472,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 472,
+ 733
+ ],
+ "score": 1.0,
+ "content": "short-term and medium duration tasks. A more detailed description of this process follows.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 45.5,
+ "bbox_fs": [
+ 105,
+ 664,
+ 506,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 170
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": "Visual MPC requires start and goal locations of objects to plan. When used in conjunction with our",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 105
+ ],
+ "score": 1.0,
+ "content": "method, we first compute the locations of objects at the bidirectional predictions and feed this in",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "place of the final goal object locations, so that Visual MPC may plan to first reach the bidirectional",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 128
+ ],
+ "score": 1.0,
+ "content": "prediction as a subgoal/stepping stone towards the final goal. Once the subgoal is reached, Visual",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 140
+ ],
+ "score": 1.0,
+ "content": "MPC is fed the final goal. To compute subgoal object locations, we use the same technique as in",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 136,
+ 505,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 136,
+ 505,
+ 151
+ ],
+ "score": 1.0,
+ "content": "Appendix F above—one-hot maps of object locations are transformed by the appearance flow maps",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 149,
+ 505,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 149,
+ 444,
+ 162
+ ],
+ "score": 1.0,
+ "content": "and masks computed internally by our predictor. This produces a probability map",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 444,
+ 150,
+ 456,
+ 161
+ ],
+ "score": 0.85,
+ "content": "p _ { g }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 456,
+ 149,
+ 505,
+ 162
+ ],
+ "score": 1.0,
+ "content": "over object",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 159,
+ 344,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 159,
+ 344,
+ 171
+ ],
+ "score": 1.0,
+ "content": "locations at the prediction, which is passed to Visual MPC.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 3.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 176,
+ 505,
+ 264
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 176,
+ 505,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 176,
+ 505,
+ 188
+ ],
+ "score": 1.0,
+ "content": "Internally, the Visual MPC planner makes fixed-time forward predictions for the object locations",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 187,
+ 504,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 249,
+ 200
+ ],
+ "score": 1.0,
+ "content": "starting from the initial distribution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 250,
+ 189,
+ 260,
+ 199
+ ],
+ "score": 0.86,
+ "content": "p _ { s }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 261,
+ 187,
+ 473,
+ 200
+ ],
+ "score": 1.0,
+ "content": "given a randomly sampled action sequence of length",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 474,
+ 187,
+ 504,
+ 198
+ ],
+ "score": 0.89,
+ "content": "h = 1 5",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 198,
+ 505,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 198,
+ 505,
+ 210
+ ],
+ "score": 1.0,
+ "content": "(the “planning horizon”). Out of all the action sequences, it selects the sequence that brings the",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 209,
+ 506,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 209,
+ 194,
+ 222
+ ],
+ "score": 1.0,
+ "content": "distribution closest to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 195,
+ 211,
+ 206,
+ 221
+ ],
+ "score": 0.86,
+ "content": "p _ { g }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 206,
+ 209,
+ 235,
+ 222
+ ],
+ "score": 1.0,
+ "content": "within",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 209,
+ 266,
+ 219
+ ],
+ "score": 0.9,
+ "content": "h = 1 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 266,
+ 209,
+ 292,
+ 222
+ ],
+ "score": 1.0,
+ "content": "steps",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 293,
+ 210,
+ 299,
+ 219
+ ],
+ "score": 0.72,
+ "content": "h",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 300,
+ 209,
+ 506,
+ 222
+ ],
+ "score": 1.0,
+ "content": "is the “planning horizon”). In practice, we use 200",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 220,
+ 505,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 505,
+ 232
+ ],
+ "score": 1.0,
+ "content": "random action sequences and perform three iterations of CEM search (Kroese et al., 2013) to find the",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 232,
+ 505,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 232,
+ 505,
+ 243
+ ],
+ "score": 1.0,
+ "content": "the best sequence. Once an action sequence is selected, the first action in the sequence is executed,",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 242,
+ 505,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 242,
+ 505,
+ 254
+ ],
+ "score": 1.0,
+ "content": "and the initial object distribution is updated. New candidate action sequences are now generated",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 253,
+ 384,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 253,
+ 384,
+ 266
+ ],
+ "score": 1.0,
+ "content": "starting from this updated object distribution and the process repeats.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 11.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 270,
+ 505,
+ 336
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 269,
+ 505,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 269,
+ 268,
+ 282
+ ],
+ "score": 1.0,
+ "content": "In our experiments, we use a time budget",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 268,
+ 270,
+ 278,
+ 280
+ ],
+ "score": 0.81,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 278,
+ 269,
+ 505,
+ 282
+ ],
+ "score": 1.0,
+ "content": "of 40 steps for 2-object pushing, and 75 steps for 3-object",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 281,
+ 506,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 281,
+ 506,
+ 293
+ ],
+ "score": 1.0,
+ "content": "pushing. In both cases, we feed Visual MPC the subgoal (bidirectional prediction object locations)",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 291,
+ 506,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 153,
+ 304
+ ],
+ "score": 1.0,
+ "content": "for the first",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 154,
+ 291,
+ 173,
+ 303
+ ],
+ "score": 0.88,
+ "content": "B / 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 173,
+ 291,
+ 408,
+ 304
+ ],
+ "score": 1.0,
+ "content": "timesteps, and then feed the final goal location for the last",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 408,
+ 291,
+ 428,
+ 303
+ ],
+ "score": 0.89,
+ "content": "B / 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 428,
+ 291,
+ 506,
+ 304
+ ],
+ "score": 1.0,
+ "content": "steps. The “direct”",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 303,
+ 505,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 303,
+ 361,
+ 315
+ ],
+ "score": 1.0,
+ "content": "planning baseline instead receives the final goal location for all",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 361,
+ 303,
+ 370,
+ 313
+ ],
+ "score": 0.8,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 371,
+ 303,
+ 505,
+ 315
+ ],
+ "score": 1.0,
+ "content": "steps. As shown by the results in",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 313,
+ 505,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 313,
+ 505,
+ 326
+ ],
+ "score": 1.0,
+ "content": "the paper, our subgoal predictions aid in performing these multi-stage pushing tasks better than with",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 325,
+ 220,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 325,
+ 220,
+ 336
+ ],
+ "score": 1.0,
+ "content": "the direct planning baseline.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 18.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 352,
+ 345,
+ 365
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 346,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 346,
+ 367
+ ],
+ "score": 1.0,
+ "content": "H SUPPLEMENTARY PREDICTION EXAMPLES",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 377,
+ 504,
+ 399
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 377,
+ 505,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 377,
+ 505,
+ 389
+ ],
+ "score": 1.0,
+ "content": "We now show more prediction results from various settings in Figs 13, 14, 15, 16, and 17, to",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 388,
+ 216,
+ 400
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 388,
+ 216,
+ 400
+ ],
+ "score": 1.0,
+ "content": "supplement those in Sec 4.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23.5
+ }
+ ],
+ "page_idx": 14,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 310,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 14,
+ "width": 13
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 170
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": "Visual MPC requires start and goal locations of objects to plan. When used in conjunction with our",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 105
+ ],
+ "score": 1.0,
+ "content": "method, we first compute the locations of objects at the bidirectional predictions and feed this in",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 505,
+ 117
+ ],
+ "score": 1.0,
+ "content": "place of the final goal object locations, so that Visual MPC may plan to first reach the bidirectional",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 128
+ ],
+ "score": 1.0,
+ "content": "prediction as a subgoal/stepping stone towards the final goal. Once the subgoal is reached, Visual",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 506,
+ 140
+ ],
+ "score": 1.0,
+ "content": "MPC is fed the final goal. To compute subgoal object locations, we use the same technique as in",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 136,
+ 505,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 136,
+ 505,
+ 151
+ ],
+ "score": 1.0,
+ "content": "Appendix F above—one-hot maps of object locations are transformed by the appearance flow maps",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 149,
+ 505,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 149,
+ 444,
+ 162
+ ],
+ "score": 1.0,
+ "content": "and masks computed internally by our predictor. This produces a probability map",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 444,
+ 150,
+ 456,
+ 161
+ ],
+ "score": 0.85,
+ "content": "p _ { g }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 456,
+ 149,
+ 505,
+ 162
+ ],
+ "score": 1.0,
+ "content": "over object",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 159,
+ 344,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 159,
+ 344,
+ 171
+ ],
+ "score": 1.0,
+ "content": "locations at the prediction, which is passed to Visual MPC.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 3.5,
+ "bbox_fs": [
+ 105,
+ 82,
+ 506,
+ 171
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 176,
+ 505,
+ 264
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 176,
+ 505,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 176,
+ 505,
+ 188
+ ],
+ "score": 1.0,
+ "content": "Internally, the Visual MPC planner makes fixed-time forward predictions for the object locations",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 187,
+ 504,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 249,
+ 200
+ ],
+ "score": 1.0,
+ "content": "starting from the initial distribution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 250,
+ 189,
+ 260,
+ 199
+ ],
+ "score": 0.86,
+ "content": "p _ { s }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 261,
+ 187,
+ 473,
+ 200
+ ],
+ "score": 1.0,
+ "content": "given a randomly sampled action sequence of length",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 474,
+ 187,
+ 504,
+ 198
+ ],
+ "score": 0.89,
+ "content": "h = 1 5",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 198,
+ 505,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 198,
+ 505,
+ 210
+ ],
+ "score": 1.0,
+ "content": "(the “planning horizon”). Out of all the action sequences, it selects the sequence that brings the",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 209,
+ 506,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 209,
+ 194,
+ 222
+ ],
+ "score": 1.0,
+ "content": "distribution closest to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 195,
+ 211,
+ 206,
+ 221
+ ],
+ "score": 0.86,
+ "content": "p _ { g }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 206,
+ 209,
+ 235,
+ 222
+ ],
+ "score": 1.0,
+ "content": "within",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 209,
+ 266,
+ 219
+ ],
+ "score": 0.9,
+ "content": "h = 1 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 266,
+ 209,
+ 292,
+ 222
+ ],
+ "score": 1.0,
+ "content": "steps",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 293,
+ 210,
+ 299,
+ 219
+ ],
+ "score": 0.72,
+ "content": "h",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 300,
+ 209,
+ 506,
+ 222
+ ],
+ "score": 1.0,
+ "content": "is the “planning horizon”). In practice, we use 200",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 220,
+ 505,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 505,
+ 232
+ ],
+ "score": 1.0,
+ "content": "random action sequences and perform three iterations of CEM search (Kroese et al., 2013) to find the",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 232,
+ 505,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 232,
+ 505,
+ 243
+ ],
+ "score": 1.0,
+ "content": "the best sequence. Once an action sequence is selected, the first action in the sequence is executed,",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 242,
+ 505,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 242,
+ 505,
+ 254
+ ],
+ "score": 1.0,
+ "content": "and the initial object distribution is updated. New candidate action sequences are now generated",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 253,
+ 384,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 253,
+ 384,
+ 266
+ ],
+ "score": 1.0,
+ "content": "starting from this updated object distribution and the process repeats.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 11.5,
+ "bbox_fs": [
+ 105,
+ 176,
+ 506,
+ 266
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 270,
+ 505,
+ 336
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 269,
+ 505,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 269,
+ 268,
+ 282
+ ],
+ "score": 1.0,
+ "content": "In our experiments, we use a time budget",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 268,
+ 270,
+ 278,
+ 280
+ ],
+ "score": 0.81,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 278,
+ 269,
+ 505,
+ 282
+ ],
+ "score": 1.0,
+ "content": "of 40 steps for 2-object pushing, and 75 steps for 3-object",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 281,
+ 506,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 281,
+ 506,
+ 293
+ ],
+ "score": 1.0,
+ "content": "pushing. In both cases, we feed Visual MPC the subgoal (bidirectional prediction object locations)",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 291,
+ 506,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 153,
+ 304
+ ],
+ "score": 1.0,
+ "content": "for the first",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 154,
+ 291,
+ 173,
+ 303
+ ],
+ "score": 0.88,
+ "content": "B / 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 173,
+ 291,
+ 408,
+ 304
+ ],
+ "score": 1.0,
+ "content": "timesteps, and then feed the final goal location for the last",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 408,
+ 291,
+ 428,
+ 303
+ ],
+ "score": 0.89,
+ "content": "B / 2",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 428,
+ 291,
+ 506,
+ 304
+ ],
+ "score": 1.0,
+ "content": "steps. The “direct”",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 303,
+ 505,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 303,
+ 361,
+ 315
+ ],
+ "score": 1.0,
+ "content": "planning baseline instead receives the final goal location for all",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 361,
+ 303,
+ 370,
+ 313
+ ],
+ "score": 0.8,
+ "content": "B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 371,
+ 303,
+ 505,
+ 315
+ ],
+ "score": 1.0,
+ "content": "steps. As shown by the results in",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 313,
+ 505,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 313,
+ 505,
+ 326
+ ],
+ "score": 1.0,
+ "content": "the paper, our subgoal predictions aid in performing these multi-stage pushing tasks better than with",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 325,
+ 220,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 325,
+ 220,
+ 336
+ ],
+ "score": 1.0,
+ "content": "the direct planning baseline.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 18.5,
+ "bbox_fs": [
+ 105,
+ 269,
+ 506,
+ 336
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 352,
+ 345,
+ 365
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 346,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 351,
+ 346,
+ 367
+ ],
+ "score": 1.0,
+ "content": "H SUPPLEMENTARY PREDICTION EXAMPLES",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 377,
+ 504,
+ 399
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 377,
+ 505,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 377,
+ 505,
+ 389
+ ],
+ "score": 1.0,
+ "content": "We now show more prediction results from various settings in Figs 13, 14, 15, 16, and 17, to",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 388,
+ 216,
+ 400
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 388,
+ 216,
+ 400
+ ],
+ "score": 1.0,
+ "content": "supplement those in Sec 4.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23.5,
+ "bbox_fs": [
+ 105,
+ 377,
+ 505,
+ 400
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 144,
+ 133,
+ 464,
+ 650
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 144,
+ 133,
+ 464,
+ 650
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 144,
+ 133,
+ 464,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 144,
+ 133,
+ 464,
+ 650
+ ],
+ "score": 0.96,
+ "type": "image",
+ "image_path": "331e0f729cf40eff5b382bbf06438a664d492baeae26a68992e41a0ef5ac3d03.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 144,
+ 133,
+ 464,
+ 305.33333333333337
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 144,
+ 305.33333333333337,
+ 464,
+ 477.66666666666674
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 144,
+ 477.66666666666674,
+ 464,
+ 650.0000000000001
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 696
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "Figure 13: (Same format as Fig 6). Supplementary bidirectional prediction results comparing fixed-time",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 677
+ ],
+ "score": 1.0,
+ "content": "prediction and our approach on grasping. First two columns are inputs (start and goal). Thereafter, each column",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 675,
+ 505,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 675,
+ 505,
+ 687
+ ],
+ "score": 1.0,
+ "content": "corresponds to the output of a different model per the column title. “match” is the ground truth image closest to",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 685,
+ 194,
+ 697
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 685,
+ 194,
+ 697
+ ],
+ "score": 1.0,
+ "content": "the GENMIN prediction.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5
+ }
+ ],
+ "index": 2.75
+ }
+ ],
+ "page_idx": 15,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "score": 1.0,
+ "content": "16",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 144,
+ 133,
+ 464,
+ 650
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 144,
+ 133,
+ 464,
+ 650
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 144,
+ 133,
+ 464,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 144,
+ 133,
+ 464,
+ 650
+ ],
+ "score": 0.96,
+ "type": "image",
+ "image_path": "331e0f729cf40eff5b382bbf06438a664d492baeae26a68992e41a0ef5ac3d03.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 144,
+ 133,
+ 464,
+ 305.33333333333337
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 144,
+ 305.33333333333337,
+ 464,
+ 477.66666666666674
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 144,
+ 477.66666666666674,
+ 464,
+ 650.0000000000001
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 696
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "Figure 13: (Same format as Fig 6). Supplementary bidirectional prediction results comparing fixed-time",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 677
+ ],
+ "score": 1.0,
+ "content": "prediction and our approach on grasping. First two columns are inputs (start and goal). Thereafter, each column",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 675,
+ 505,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 675,
+ 505,
+ 687
+ ],
+ "score": 1.0,
+ "content": "corresponds to the output of a different model per the column title. “match” is the ground truth image closest to",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 685,
+ 194,
+ 697
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 685,
+ 194,
+ 697
+ ],
+ "score": 1.0,
+ "content": "the GENMIN prediction.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5
+ }
+ ],
+ "index": 2.75
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 144,
+ 131,
+ 464,
+ 650
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 144,
+ 131,
+ 464,
+ 650
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 144,
+ 131,
+ 464,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 144,
+ 131,
+ 464,
+ 650
+ ],
+ "score": 0.958,
+ "type": "image",
+ "image_path": "4c02c50243df3f49393e35f68fcbe9a306dde711b7ab7db4fa781dd6ed57197c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 144,
+ 131,
+ 464,
+ 304.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 144,
+ 304.0,
+ 464,
+ 477.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 144,
+ 477.0,
+ 464,
+ 650.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 696
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 506,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 506,
+ 668
+ ],
+ "score": 1.0,
+ "content": "Figure 14: (Same format as Fig 6). Supplementary bidirectional prediction results comparing fixed-time",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 677
+ ],
+ "score": 1.0,
+ "content": "prediction and our approach on pick-and-place. First two columns are inputs (start and goal). Thereafter, each",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 674,
+ 505,
+ 688
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 674,
+ 505,
+ 688
+ ],
+ "score": 1.0,
+ "content": "column corresponds to the output of a different model per the column title. “match” is the ground truth image",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 685,
+ 230,
+ 697
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 685,
+ 230,
+ 697
+ ],
+ "score": 1.0,
+ "content": "closest to the GENMIN prediction.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5
+ }
+ ],
+ "index": 2.75
+ }
+ ],
+ "page_idx": 16,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 310,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 14,
+ "width": 13
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 144,
+ 131,
+ 464,
+ 650
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 144,
+ 131,
+ 464,
+ 650
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 144,
+ 131,
+ 464,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 144,
+ 131,
+ 464,
+ 650
+ ],
+ "score": 0.958,
+ "type": "image",
+ "image_path": "4c02c50243df3f49393e35f68fcbe9a306dde711b7ab7db4fa781dd6ed57197c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 144,
+ 131,
+ 464,
+ 304.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 144,
+ 304.0,
+ 464,
+ 477.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 144,
+ 477.0,
+ 464,
+ 650.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 696
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 506,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 506,
+ 668
+ ],
+ "score": 1.0,
+ "content": "Figure 14: (Same format as Fig 6). Supplementary bidirectional prediction results comparing fixed-time",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 677
+ ],
+ "score": 1.0,
+ "content": "prediction and our approach on pick-and-place. First two columns are inputs (start and goal). Thereafter, each",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 674,
+ 505,
+ 688
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 674,
+ 505,
+ 688
+ ],
+ "score": 1.0,
+ "content": "column corresponds to the output of a different model per the column title. “match” is the ground truth image",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 685,
+ 230,
+ 697
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 685,
+ 230,
+ 697
+ ],
+ "score": 1.0,
+ "content": "closest to the GENMIN prediction.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5
+ }
+ ],
+ "index": 2.75
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 0,
+ 168,
+ 595,
+ 613
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 0,
+ 168,
+ 595,
+ 613
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 0,
+ 168,
+ 595,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 0,
+ 168,
+ 595,
+ 613
+ ],
+ "score": 0.965,
+ "type": "image",
+ "image_path": "7608e21d6351de86eb9561103060c7e34e988b183e26ac97471c7f928e861972.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 0,
+ 168,
+ 595,
+ 316.33333333333337
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 0,
+ 316.33333333333337,
+ 595,
+ 464.66666666666674
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 0,
+ 464.66666666666674,
+ 595,
+ 613.0000000000001
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 618,
+ 504,
+ 650
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 618,
+ 505,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 618,
+ 505,
+ 630
+ ],
+ "score": 1.0,
+ "content": "Figure 15: Supplementary bidirectional prediction results on pushing. Four columns for each method correspond",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 629,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 629,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "to masked warped start frame, masked warped goal frame, masked new pixels, and final output respectively. See",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 639,
+ 366,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 639,
+ 366,
+ 651
+ ],
+ "score": 1.0,
+ "content": "Appendix C and Fig 12 for the synthesis scheme with masks and warps.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ }
+ ],
+ "page_idx": 17,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 38
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 763
+ ],
+ "score": 1.0,
+ "content": "18",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 0,
+ 168,
+ 595,
+ 613
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 0,
+ 168,
+ 595,
+ 613
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 0,
+ 168,
+ 595,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 0,
+ 168,
+ 595,
+ 613
+ ],
+ "score": 0.965,
+ "type": "image",
+ "image_path": "7608e21d6351de86eb9561103060c7e34e988b183e26ac97471c7f928e861972.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 0,
+ 168,
+ 595,
+ 316.33333333333337
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 0,
+ 316.33333333333337,
+ 595,
+ 464.66666666666674
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 0,
+ 464.66666666666674,
+ 595,
+ 613.0000000000001
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 618,
+ 504,
+ 650
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 618,
+ 505,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 618,
+ 505,
+ 630
+ ],
+ "score": 1.0,
+ "content": "Figure 15: Supplementary bidirectional prediction results on pushing. Four columns for each method correspond",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 629,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 629,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "to masked warped start frame, masked warped goal frame, masked new pixels, and final output respectively. See",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 639,
+ 366,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 639,
+ 366,
+ 651
+ ],
+ "score": 1.0,
+ "content": "Appendix C and Fig 12 for the synthesis scheme with masks and warps.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 41,
+ 166,
+ 558,
+ 612
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 41,
+ 166,
+ 558,
+ 612
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 41,
+ 166,
+ 558,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 41,
+ 166,
+ 558,
+ 612
+ ],
+ "score": 0.967,
+ "type": "image",
+ "image_path": "ef3dab5ffbd647668dd9b6a11921ad8c76166914d6c8c768742020a369945769.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 41,
+ 166,
+ 558,
+ 314.66666666666663
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 41,
+ 314.66666666666663,
+ 558,
+ 463.33333333333326
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 41,
+ 463.33333333333326,
+ 558,
+ 611.9999999999999
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 618,
+ 506,
+ 660
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 619,
+ 505,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 619,
+ 455,
+ 631
+ ],
+ "score": 1.0,
+ "content": "Figure 16: Supplementary bidirectional prediction examples for pick-and-place with GENMIN",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 455,
+ 621,
+ 461,
+ 628
+ ],
+ "score": 0.42,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 462,
+ 619,
+ 505,
+ 631
+ ],
+ "score": 1.0,
+ "content": "VAE. Same",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 628,
+ 505,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 628,
+ 417,
+ 641
+ ],
+ "score": 1.0,
+ "content": "format as Fig 9. Each row is a separate example. First column is the input. GENMIN",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 417,
+ 630,
+ 424,
+ 637
+ ],
+ "score": 0.37,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 424,
+ 628,
+ 505,
+ 641
+ ],
+ "score": 1.0,
+ "content": "VAE captures residual",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 639,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 639,
+ 250,
+ 650
+ ],
+ "score": 1.0,
+ "content": "stochasticity at the bottleneck. GENMIN",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 250,
+ 640,
+ 257,
+ 648
+ ],
+ "score": 0.43,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 257,
+ 639,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "VAE produces images that are most all of the arm in contact with the",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 442,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 442,
+ 661
+ ],
+ "score": 1.0,
+ "content": "object on the table, but at different points on the object, and with different arm/gripper poses.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5
+ }
+ ],
+ "index": 2.75
+ }
+ ],
+ "page_idx": 18,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 312,
+ 764
+ ],
+ "score": 1.0,
+ "content": "19",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 41,
+ 166,
+ 558,
+ 612
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 41,
+ 166,
+ 558,
+ 612
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 41,
+ 166,
+ 558,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 41,
+ 166,
+ 558,
+ 612
+ ],
+ "score": 0.967,
+ "type": "image",
+ "image_path": "ef3dab5ffbd647668dd9b6a11921ad8c76166914d6c8c768742020a369945769.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 41,
+ 166,
+ 558,
+ 314.66666666666663
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 41,
+ 314.66666666666663,
+ 558,
+ 463.33333333333326
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 41,
+ 463.33333333333326,
+ 558,
+ 611.9999999999999
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 618,
+ 506,
+ 660
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 619,
+ 505,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 619,
+ 455,
+ 631
+ ],
+ "score": 1.0,
+ "content": "Figure 16: Supplementary bidirectional prediction examples for pick-and-place with GENMIN",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 455,
+ 621,
+ 461,
+ 628
+ ],
+ "score": 0.42,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 462,
+ 619,
+ 505,
+ 631
+ ],
+ "score": 1.0,
+ "content": "VAE. Same",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 628,
+ 505,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 628,
+ 417,
+ 641
+ ],
+ "score": 1.0,
+ "content": "format as Fig 9. Each row is a separate example. First column is the input. GENMIN",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 417,
+ 630,
+ 424,
+ 637
+ ],
+ "score": 0.37,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 424,
+ 628,
+ 505,
+ 641
+ ],
+ "score": 1.0,
+ "content": "VAE captures residual",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 639,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 639,
+ 250,
+ 650
+ ],
+ "score": 1.0,
+ "content": "stochasticity at the bottleneck. GENMIN",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 250,
+ 640,
+ 257,
+ 648
+ ],
+ "score": 0.43,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 257,
+ 639,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "VAE produces images that are most all of the arm in contact with the",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 442,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 442,
+ 661
+ ],
+ "score": 1.0,
+ "content": "object on the table, but at different points on the object, and with different arm/gripper poses.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5
+ }
+ ],
+ "index": 2.75
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 42,
+ 157,
+ 559,
+ 631
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 42,
+ 157,
+ 559,
+ 631
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 42,
+ 157,
+ 559,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 42,
+ 157,
+ 559,
+ 631
+ ],
+ "score": 0.965,
+ "type": "image",
+ "image_path": "935acf73750df876dda1f586688730675d4fa8665ac4a901053f00f014cd5c71.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 42,
+ 157,
+ 559,
+ 315.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 42,
+ 315.0,
+ 559,
+ 473.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 42,
+ 473.0,
+ 559,
+ 631.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 636,
+ 505,
+ 668
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 649
+ ],
+ "score": 1.0,
+ "content": "Figure 17: Supplementary forward prediction examples for grasping, comparing fixed-time predictors and our",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 646,
+ 506,
+ 659
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 646,
+ 506,
+ 659
+ ],
+ "score": 1.0,
+ "content": "approach. Same format as Fig 5. Each row is a separate example. First column is the input. Thereafter, each",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 657,
+ 376,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 657,
+ 376,
+ 668
+ ],
+ "score": 1.0,
+ "content": "column corresponds to the output of a different model per the column title.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ }
+ ],
+ "page_idx": 19,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 293,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2019",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 298,
+ 749,
+ 313,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 298,
+ 749,
+ 313,
+ 764
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 15,
+ "width": 15
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 42,
+ 157,
+ 559,
+ 631
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 42,
+ 157,
+ 559,
+ 631
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 42,
+ 157,
+ 559,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 42,
+ 157,
+ 559,
+ 631
+ ],
+ "score": 0.965,
+ "type": "image",
+ "image_path": "935acf73750df876dda1f586688730675d4fa8665ac4a901053f00f014cd5c71.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 42,
+ 157,
+ 559,
+ 315.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 42,
+ 315.0,
+ 559,
+ 473.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 42,
+ 473.0,
+ 559,
+ 631.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 636,
+ 505,
+ 668
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 649
+ ],
+ "score": 1.0,
+ "content": "Figure 17: Supplementary forward prediction examples for grasping, comparing fixed-time predictors and our",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 646,
+ 506,
+ 659
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 646,
+ 506,
+ 659
+ ],
+ "score": 1.0,
+ "content": "approach. Same format as Fig 5. Each row is a separate example. First column is the input. Thereafter, each",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 657,
+ 376,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 657,
+ 376,
+ 668
+ ],
+ "score": 1.0,
+ "content": "column corresponds to the output of a different model per the column title.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ }
+ ]
+ }
+ ],
+ "_backend": "pipeline",
+ "_version_name": "2.2.2"
+}
\ No newline at end of file