diff --git "a/parse/train/HJgpugrKPS/HJgpugrKPS_middle.json" "b/parse/train/HJgpugrKPS/HJgpugrKPS_middle.json"
new file mode 100644--- /dev/null
+++ "b/parse/train/HJgpugrKPS/HJgpugrKPS_middle.json"
@@ -0,0 +1,33862 @@
+{
+ "pdf_info": [
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 78,
+ 450,
+ 96
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 77,
+ 453,
+ 101
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 77,
+ 453,
+ 101
+ ],
+ "score": 1.0,
+ "content": "SCALE-EQUIVARIANT STEERABLE NETWORKS",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 112,
+ 116,
+ 339,
+ 150
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 117,
+ 339,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 117,
+ 339,
+ 129
+ ],
+ "score": 1.0,
+ "content": "Ivan Sosnovik∗, Michał Szmaja†, Arnold Smeulders",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 113,
+ 128,
+ 204,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 128,
+ 204,
+ 139
+ ],
+ "score": 1.0,
+ "content": "UvA-Bosch Delta Lab",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 112,
+ 140,
+ 216,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 140,
+ 216,
+ 151
+ ],
+ "score": 1.0,
+ "content": "University of Amsterdam",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 180,
+ 333,
+ 192
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 179,
+ 335,
+ 193
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 179,
+ 335,
+ 193
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 205,
+ 468,
+ 337
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 205,
+ 468,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 205,
+ 468,
+ 218
+ ],
+ "score": 1.0,
+ "content": "The effectiveness of Convolutional Neural Networks (CNNs) has been substan-",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 141,
+ 216,
+ 469,
+ 230
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 216,
+ 469,
+ 230
+ ],
+ "score": 1.0,
+ "content": "tially attributed to their built-in property of translation equivariance. However,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 141,
+ 227,
+ 469,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 227,
+ 469,
+ 240
+ ],
+ "score": 1.0,
+ "content": "CNNs do not have embedded mechanisms to handle other types of transforma-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 141,
+ 238,
+ 469,
+ 251
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 238,
+ 469,
+ 251
+ ],
+ "score": 1.0,
+ "content": "tions. In this work, we pay attention to scale changes, which regularly appear in",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 142,
+ 250,
+ 469,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 250,
+ 469,
+ 261
+ ],
+ "score": 1.0,
+ "content": "various tasks due to the changing distances between the objects and the camera.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 260,
+ 469,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 260,
+ 469,
+ 273
+ ],
+ "score": 1.0,
+ "content": "First, we introduce the general theory for building scale-equivariant convolutional",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 141,
+ 272,
+ 469,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 272,
+ 469,
+ 283
+ ],
+ "score": 1.0,
+ "content": "networks with steerable filters. We develop scale-convolution and generalize other",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 141,
+ 282,
+ 470,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 282,
+ 470,
+ 295
+ ],
+ "score": 1.0,
+ "content": "common blocks to be scale-equivariant. We demonstrate the computational effi-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 293,
+ 470,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 293,
+ 470,
+ 306
+ ],
+ "score": 1.0,
+ "content": "ciency and numerical stability of the proposed method. We compare the proposed",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 304,
+ 469,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 304,
+ 469,
+ 316
+ ],
+ "score": 1.0,
+ "content": "models to the previously developed methods for scale equivariance and local scale",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 142,
+ 316,
+ 470,
+ 327
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 316,
+ 470,
+ 327
+ ],
+ "score": 1.0,
+ "content": "invariance. We demonstrate state-of-the-art results on the MNIST-scale dataset",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 141,
+ 325,
+ 387,
+ 340
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 325,
+ 387,
+ 340
+ ],
+ "score": 1.0,
+ "content": "and on the STL-10 dataset in the supervised learning setting.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 10.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 360,
+ 206,
+ 372
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 359,
+ 208,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 359,
+ 208,
+ 376
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 385,
+ 505,
+ 518
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 386,
+ 505,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 386,
+ 505,
+ 398
+ ],
+ "score": 1.0,
+ "content": "Scale transformations occur in many image and video analysis tasks. They are a natural consequence",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 409
+ ],
+ "score": 1.0,
+ "content": "of the variable distances among objects, or between objects and the camera. Such transformations",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 421
+ ],
+ "score": 1.0,
+ "content": "result in significant changes in the input space which are often difficult for models to handle ap-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 505,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 505,
+ 432
+ ],
+ "score": 1.0,
+ "content": "propriately without careful consideration. At a high level, there are two modeling paradigms which",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 430,
+ 505,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 430,
+ 505,
+ 442
+ ],
+ "score": 1.0,
+ "content": "allow a model to deal with scale changes: models can be endowed with an internal notion of scale",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 441,
+ 505,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 441,
+ 505,
+ 453
+ ],
+ "score": 1.0,
+ "content": "and transform their predictions accordingly, or instead, models can be designed to be specifically",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 452,
+ 505,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 452,
+ 505,
+ 463
+ ],
+ "score": 1.0,
+ "content": "invariant to scale changes. In image classification, when scale changes are commonly a factor of 2,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 461,
+ 505,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 461,
+ 505,
+ 475
+ ],
+ "score": 1.0,
+ "content": "it is often sufficient to make class prediction independent of scale. However, in tasks such as image",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 474,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 474,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "segmentation, visual tracking, or object detection, scale changes can reach factors of 10 or more.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 484,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 484,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "In these cases, it is intuitive that the ideal prediction should scale proportionally to the input. For",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 496,
+ 505,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 496,
+ 505,
+ 508
+ ],
+ "score": 1.0,
+ "content": "example, the segmentation map of a nearby pedestrian should be easily converted to that of a distant",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 506,
+ 233,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 506,
+ 233,
+ 520
+ ],
+ "score": 1.0,
+ "content": "person simply by downscaling.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 23.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 523,
+ 505,
+ 600
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 521,
+ 505,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 521,
+ 505,
+ 537
+ ],
+ "score": 1.0,
+ "content": "Convolutional Neural Networks (CNNs) demonstrate state-of-the-art performance in a wide range",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 535,
+ 505,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 535,
+ 505,
+ 546
+ ],
+ "score": 1.0,
+ "content": "of tasks. Yet, despite their built-in translation equivariance, they do not have a particular mechanism",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 546,
+ 505,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 546,
+ 505,
+ 557
+ ],
+ "score": 1.0,
+ "content": "for dealing with scale changes. One way to make CNNs account for scale is to train them with data",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 104,
+ 555,
+ 505,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 555,
+ 505,
+ 569
+ ],
+ "score": 1.0,
+ "content": "augmentation Barnard & Casasent (1991). This is, however, suitable only for global transformations.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 567,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 567,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": "As an alternative, Henriques & Vedaldi (2017) and Tai et al. (2019) use the canonical coordinates",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 578,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 578,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": "of scale transformations to reduce scaling to well-studied translations. While these approaches do",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 589,
+ 422,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 589,
+ 422,
+ 602
+ ],
+ "score": 1.0,
+ "content": "allow for scale equivariance, they consequently break translation equivariance.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 605,
+ 505,
+ 694
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "Several attempts have thus been made to extend CNNs to both scale and translation symmetry si-",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 629
+ ],
+ "score": 1.0,
+ "content": "multaneously. Some works use input or filter resizing to account for scaling in deep layers Xu et al.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "(2014); Kanazawa et al. (2014). Such methods are suboptimal due to the time complexity of tensor",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 639,
+ 506,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 639,
+ 506,
+ 651
+ ],
+ "score": 1.0,
+ "content": "resizing and the need for interpolation. In Ghosh & Gupta (2019) the authors pre-calculate filters",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 649,
+ 506,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 506,
+ 663
+ ],
+ "score": 1.0,
+ "content": "defined on several scales to build scale-invariant networks, while ignoring the important case of",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 659,
+ 505,
+ 674
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 659,
+ 505,
+ 674
+ ],
+ "score": 1.0,
+ "content": "scale equivariance. In contrast, Worrall & Welling (2019) employ the theory of semigroup equiv-",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 672,
+ 505,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 672,
+ 505,
+ 685
+ ],
+ "score": 1.0,
+ "content": "ariant networks with scale-space as an example; however, this method is only suitable for integer",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 683,
+ 268,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 683,
+ 268,
+ 694
+ ],
+ "score": 1.0,
+ "content": "downscale factors and therefore limited.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 40.5
+ }
+ ],
+ "page_idx": 0,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 118,
+ 701,
+ 410,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 700,
+ 286,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 700,
+ 286,
+ 712
+ ],
+ "score": 1.0,
+ "content": "∗Correspondence to i.sosnovik@uva.nl",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 118,
+ 710,
+ 410,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 710,
+ 410,
+ 723
+ ],
+ "score": 1.0,
+ "content": "†Work was done when the author had an internship at the UvA-Bosch Delta Lab",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 122,
+ 721,
+ 394,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 122,
+ 721,
+ 394,
+ 732
+ ],
+ "score": 1.0,
+ "content": "Source code is available at http://github.com/isosnovik/sesn",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 294,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 303,
+ 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": [
+ 108,
+ 78,
+ 450,
+ 96
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 77,
+ 453,
+ 101
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 77,
+ 453,
+ 101
+ ],
+ "score": 1.0,
+ "content": "SCALE-EQUIVARIANT STEERABLE NETWORKS",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 112,
+ 116,
+ 339,
+ 150
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 117,
+ 339,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 117,
+ 339,
+ 129
+ ],
+ "score": 1.0,
+ "content": "Ivan Sosnovik∗, Michał Szmaja†, Arnold Smeulders",
+ "type": "text"
+ }
+ ],
+ "index": 1,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 113,
+ 128,
+ 204,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 128,
+ 204,
+ 139
+ ],
+ "score": 1.0,
+ "content": "UvA-Bosch Delta Lab",
+ "type": "text"
+ }
+ ],
+ "index": 2,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 112,
+ 140,
+ 216,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 140,
+ 216,
+ 151
+ ],
+ "score": 1.0,
+ "content": "University of Amsterdam",
+ "type": "text"
+ }
+ ],
+ "index": 3,
+ "is_list_start_line": true
+ }
+ ],
+ "index": 2,
+ "bbox_fs": [
+ 111,
+ 117,
+ 339,
+ 151
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 180,
+ 333,
+ 192
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 179,
+ 335,
+ 193
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 179,
+ 335,
+ 193
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 205,
+ 468,
+ 337
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 205,
+ 468,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 205,
+ 468,
+ 218
+ ],
+ "score": 1.0,
+ "content": "The effectiveness of Convolutional Neural Networks (CNNs) has been substan-",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 141,
+ 216,
+ 469,
+ 230
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 216,
+ 469,
+ 230
+ ],
+ "score": 1.0,
+ "content": "tially attributed to their built-in property of translation equivariance. However,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 141,
+ 227,
+ 469,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 227,
+ 469,
+ 240
+ ],
+ "score": 1.0,
+ "content": "CNNs do not have embedded mechanisms to handle other types of transforma-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 141,
+ 238,
+ 469,
+ 251
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 238,
+ 469,
+ 251
+ ],
+ "score": 1.0,
+ "content": "tions. In this work, we pay attention to scale changes, which regularly appear in",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 142,
+ 250,
+ 469,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 250,
+ 469,
+ 261
+ ],
+ "score": 1.0,
+ "content": "various tasks due to the changing distances between the objects and the camera.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 260,
+ 469,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 260,
+ 469,
+ 273
+ ],
+ "score": 1.0,
+ "content": "First, we introduce the general theory for building scale-equivariant convolutional",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 141,
+ 272,
+ 469,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 272,
+ 469,
+ 283
+ ],
+ "score": 1.0,
+ "content": "networks with steerable filters. We develop scale-convolution and generalize other",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 141,
+ 282,
+ 470,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 282,
+ 470,
+ 295
+ ],
+ "score": 1.0,
+ "content": "common blocks to be scale-equivariant. We demonstrate the computational effi-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 293,
+ 470,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 293,
+ 470,
+ 306
+ ],
+ "score": 1.0,
+ "content": "ciency and numerical stability of the proposed method. We compare the proposed",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 304,
+ 469,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 304,
+ 469,
+ 316
+ ],
+ "score": 1.0,
+ "content": "models to the previously developed methods for scale equivariance and local scale",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 142,
+ 316,
+ 470,
+ 327
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 316,
+ 470,
+ 327
+ ],
+ "score": 1.0,
+ "content": "invariance. We demonstrate state-of-the-art results on the MNIST-scale dataset",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 141,
+ 325,
+ 387,
+ 340
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 325,
+ 387,
+ 340
+ ],
+ "score": 1.0,
+ "content": "and on the STL-10 dataset in the supervised learning setting.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 10.5,
+ "bbox_fs": [
+ 141,
+ 205,
+ 470,
+ 340
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 360,
+ 206,
+ 372
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 359,
+ 208,
+ 376
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 359,
+ 208,
+ 376
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 385,
+ 505,
+ 518
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 386,
+ 505,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 386,
+ 505,
+ 398
+ ],
+ "score": 1.0,
+ "content": "Scale transformations occur in many image and video analysis tasks. They are a natural consequence",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 396,
+ 505,
+ 409
+ ],
+ "score": 1.0,
+ "content": "of the variable distances among objects, or between objects and the camera. Such transformations",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 421
+ ],
+ "score": 1.0,
+ "content": "result in significant changes in the input space which are often difficult for models to handle ap-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 505,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 505,
+ 432
+ ],
+ "score": 1.0,
+ "content": "propriately without careful consideration. At a high level, there are two modeling paradigms which",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 430,
+ 505,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 430,
+ 505,
+ 442
+ ],
+ "score": 1.0,
+ "content": "allow a model to deal with scale changes: models can be endowed with an internal notion of scale",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 441,
+ 505,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 441,
+ 505,
+ 453
+ ],
+ "score": 1.0,
+ "content": "and transform their predictions accordingly, or instead, models can be designed to be specifically",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 452,
+ 505,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 452,
+ 505,
+ 463
+ ],
+ "score": 1.0,
+ "content": "invariant to scale changes. In image classification, when scale changes are commonly a factor of 2,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 461,
+ 505,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 461,
+ 505,
+ 475
+ ],
+ "score": 1.0,
+ "content": "it is often sufficient to make class prediction independent of scale. However, in tasks such as image",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 474,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 474,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "segmentation, visual tracking, or object detection, scale changes can reach factors of 10 or more.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 484,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 484,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "In these cases, it is intuitive that the ideal prediction should scale proportionally to the input. For",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 496,
+ 505,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 496,
+ 505,
+ 508
+ ],
+ "score": 1.0,
+ "content": "example, the segmentation map of a nearby pedestrian should be easily converted to that of a distant",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 506,
+ 233,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 506,
+ 233,
+ 520
+ ],
+ "score": 1.0,
+ "content": "person simply by downscaling.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 23.5,
+ "bbox_fs": [
+ 105,
+ 386,
+ 505,
+ 520
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 523,
+ 505,
+ 600
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 521,
+ 505,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 521,
+ 505,
+ 537
+ ],
+ "score": 1.0,
+ "content": "Convolutional Neural Networks (CNNs) demonstrate state-of-the-art performance in a wide range",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 535,
+ 505,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 535,
+ 505,
+ 546
+ ],
+ "score": 1.0,
+ "content": "of tasks. Yet, despite their built-in translation equivariance, they do not have a particular mechanism",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 546,
+ 505,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 546,
+ 505,
+ 557
+ ],
+ "score": 1.0,
+ "content": "for dealing with scale changes. One way to make CNNs account for scale is to train them with data",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 104,
+ 555,
+ 505,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 555,
+ 505,
+ 569
+ ],
+ "score": 1.0,
+ "content": "augmentation Barnard & Casasent (1991). This is, however, suitable only for global transformations.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 567,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 567,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": "As an alternative, Henriques & Vedaldi (2017) and Tai et al. (2019) use the canonical coordinates",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 578,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 578,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": "of scale transformations to reduce scaling to well-studied translations. While these approaches do",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 589,
+ 422,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 589,
+ 422,
+ 602
+ ],
+ "score": 1.0,
+ "content": "allow for scale equivariance, they consequently break translation equivariance.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 33,
+ "bbox_fs": [
+ 104,
+ 521,
+ 505,
+ 602
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 605,
+ 505,
+ 694
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "Several attempts have thus been made to extend CNNs to both scale and translation symmetry si-",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 629
+ ],
+ "score": 1.0,
+ "content": "multaneously. Some works use input or filter resizing to account for scaling in deep layers Xu et al.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "(2014); Kanazawa et al. (2014). Such methods are suboptimal due to the time complexity of tensor",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 639,
+ 506,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 639,
+ 506,
+ 651
+ ],
+ "score": 1.0,
+ "content": "resizing and the need for interpolation. In Ghosh & Gupta (2019) the authors pre-calculate filters",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 649,
+ 506,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 506,
+ 663
+ ],
+ "score": 1.0,
+ "content": "defined on several scales to build scale-invariant networks, while ignoring the important case of",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 659,
+ 505,
+ 674
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 659,
+ 505,
+ 674
+ ],
+ "score": 1.0,
+ "content": "scale equivariance. In contrast, Worrall & Welling (2019) employ the theory of semigroup equiv-",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 672,
+ 505,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 672,
+ 505,
+ 685
+ ],
+ "score": 1.0,
+ "content": "ariant networks with scale-space as an example; however, this method is only suitable for integer",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 683,
+ 268,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 683,
+ 268,
+ 694
+ ],
+ "score": 1.0,
+ "content": "downscale factors and therefore limited.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 105,
+ 606,
+ 506,
+ 694
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 149
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": "In this paper we develop a theory of scale-equivariant networks. We demonstrate the concept of",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "steerable filter parametrization which allows for scaling without the need for tensor resizing. Then",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 105,
+ 504,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 105,
+ 504,
+ 117
+ ],
+ "score": 1.0,
+ "content": "we derive scale-equivariant convolution and demonstrate a fast algorithm for its implementation.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 128
+ ],
+ "score": 1.0,
+ "content": "Furthermore, we experiment to determine to what degree the mathematical properties actually hold",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 138
+ ],
+ "score": 1.0,
+ "content": "true. Finally, we conduct a set of experiments comparing our model with other methods for scale",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 138,
+ 267,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 138,
+ 267,
+ 149
+ ],
+ "score": 1.0,
+ "content": "equivariance and local scale invariance.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 168,
+ 487,
+ 180
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 168,
+ 488,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 168,
+ 488,
+ 182
+ ],
+ "score": 1.0,
+ "content": "The proposed model has the following advantages compared to other scale-equivariant models:",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 129,
+ 189,
+ 505,
+ 275
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 190,
+ 505,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 190,
+ 505,
+ 201
+ ],
+ "score": 1.0,
+ "content": "1. It is equivariant to scale transformations with arbitrary discrete scale factors and is not",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 141,
+ 200,
+ 431,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 200,
+ 431,
+ 213
+ ],
+ "score": 1.0,
+ "content": "limited to either integer scales or scales tailored by the image pixel grid.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 129,
+ 215,
+ 505,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 215,
+ 505,
+ 228
+ ],
+ "score": 1.0,
+ "content": "2. It does not rely on any image resampling techniques during training, and therefore, pro-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 226,
+ 457,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 226,
+ 457,
+ 239
+ ],
+ "score": 1.0,
+ "content": "duces deep scale-equivariant representations free of any interpolation artifacts.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 129,
+ 241,
+ 505,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 241,
+ 505,
+ 254
+ ],
+ "score": 1.0,
+ "content": "3. The algorithm is based on the combination of tensor expansion and 2-dimensional convo-",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 141,
+ 252,
+ 506,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 252,
+ 506,
+ 265
+ ],
+ "score": 1.0,
+ "content": "lution, and demonstrates the same computation time as the general CNN with a comparable",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 142,
+ 263,
+ 188,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 263,
+ 188,
+ 275
+ ],
+ "score": 1.0,
+ "content": "filter bank.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 290,
+ 208,
+ 303
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 289,
+ 209,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 289,
+ 209,
+ 306
+ ],
+ "score": 1.0,
+ "content": "2 PRELIMINARIES",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 315,
+ 505,
+ 360
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 505,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 505,
+ 329
+ ],
+ "score": 1.0,
+ "content": "Before we move into scale-equivariant mappings, we discuss some aspects of equivariance, scaling",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 327,
+ 504,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 327,
+ 504,
+ 339
+ ],
+ "score": 1.0,
+ "content": "transformations, symmetry groups, and the functions defined on them. For simplicity, in this sec-",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 338,
+ 505,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 338,
+ 505,
+ 350
+ ],
+ "score": 1.0,
+ "content": "tion, we consider only 1-dimensional functions. The generalization to higher-dimensional cases is",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 348,
+ 173,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 348,
+ 173,
+ 361
+ ],
+ "score": 1.0,
+ "content": "straightforward.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 16.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 365,
+ 504,
+ 388
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 365,
+ 505,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 365,
+ 289,
+ 378
+ ],
+ "score": 1.0,
+ "content": "Equivariance Let us consider some mapping",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 290,
+ 368,
+ 296,
+ 377
+ ],
+ "score": 0.74,
+ "content": "g",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 297,
+ 365,
+ 390,
+ 378
+ ],
+ "score": 1.0,
+ "content": ". It is equivariant under",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 390,
+ 366,
+ 402,
+ 376
+ ],
+ "score": 0.88,
+ "content": "L _ { \\theta }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 402,
+ 365,
+ 505,
+ 378
+ ],
+ "score": 1.0,
+ "content": "if and only if there exists",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 107,
+ 375,
+ 477,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 377,
+ 119,
+ 389
+ ],
+ "score": 0.89,
+ "content": "L _ { \\theta } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 119,
+ 375,
+ 158,
+ 390
+ ],
+ "score": 1.0,
+ "content": "such that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 159,
+ 376,
+ 225,
+ 389
+ ],
+ "score": 0.93,
+ "content": "g \\circ L _ { \\theta } = L _ { \\theta } ^ { \\prime } \\circ g",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 225,
+ 375,
+ 259,
+ 390
+ ],
+ "score": 1.0,
+ "content": ". In case",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 376,
+ 272,
+ 389
+ ],
+ "score": 0.92,
+ "content": "L _ { \\theta } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 272,
+ 375,
+ 420,
+ 390
+ ],
+ "score": 1.0,
+ "content": "is the identity mapping, the function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 420,
+ 378,
+ 426,
+ 388
+ ],
+ "score": 0.79,
+ "content": "g",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 427,
+ 375,
+ 477,
+ 390
+ ],
+ "score": 1.0,
+ "content": "is invariant.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 19.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 393,
+ 505,
+ 426
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 393,
+ 505,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 393,
+ 505,
+ 405
+ ],
+ "score": 1.0,
+ "content": "In this paper we consider scaling transformations. In order to guarantee the equivariance of the",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 416
+ ],
+ "score": 1.0,
+ "content": "predictions to such transformations, and to improve the performance of the model, we seek to incor-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 276,
+ 427
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 276,
+ 427
+ ],
+ "score": 1.0,
+ "content": "porate this property directly inside CNNs.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 432,
+ 435,
+ 444
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 431,
+ 435,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 431,
+ 209,
+ 445
+ ],
+ "score": 1.0,
+ "content": "Scaling Given a function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 209,
+ 432,
+ 254,
+ 444
+ ],
+ "score": 0.92,
+ "content": "f : \\mathbb { R } \\to \\mathbb { R }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 255,
+ 431,
+ 435,
+ 445
+ ],
+ "score": 1.0,
+ "content": ", a scale transformation is defined as follows:",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 240,
+ 448,
+ 371,
+ 463
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 240,
+ 448,
+ 371,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 240,
+ 448,
+ 371,
+ 463
+ ],
+ "score": 0.91,
+ "content": "L _ { s } [ f ] ( x ) = f ( s ^ { - 1 } x ) , \\quad \\forall s > 0",
+ "type": "interline_equation",
+ "image_path": "13908d2afa72d06c5f750fd9545ff5fd44438e6667a9eef4ad001702836f5831.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 25,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 240,
+ 448,
+ 371,
+ 463
+ ],
+ "spans": [],
+ "index": 25
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 474,
+ 505,
+ 507
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 474,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 474,
+ 199,
+ 486
+ ],
+ "score": 1.0,
+ "content": "We refer to cases with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 199,
+ 475,
+ 224,
+ 484
+ ],
+ "score": 0.9,
+ "content": "s > 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 224,
+ 474,
+ 341,
+ 486
+ ],
+ "score": 1.0,
+ "content": "as upscale and to cases with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 342,
+ 475,
+ 367,
+ 484
+ ],
+ "score": 0.9,
+ "content": "s < 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 367,
+ 474,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "as downscale. If we convolve the",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 485,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 485,
+ 284,
+ 497
+ ],
+ "score": 1.0,
+ "content": "downscaled function with an arbitrary filter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 284,
+ 486,
+ 292,
+ 497
+ ],
+ "score": 0.85,
+ "content": "\\psi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 293,
+ 485,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "and perform a simple change of variables inside the",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 495,
+ 265,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 265,
+ 509
+ ],
+ "score": 1.0,
+ "content": "integral, we get the following property:",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 131,
+ 512,
+ 480,
+ 567
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 512,
+ 480,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 512,
+ 480,
+ 567
+ ],
+ "score": 0.95,
+ "content": "\\begin{array} { l } { \\displaystyle [ L _ { s } [ f ] \\star \\psi ] ( x ) = \\int _ { \\mathbb R } L _ { s } [ f ] ( x ^ { \\prime } ) \\psi ( x ^ { \\prime } - x ) d x ^ { \\prime } = \\int _ { \\mathbb R } f ( s ^ { - 1 } x ^ { \\prime } ) \\psi ( x ^ { \\prime } - x ) d x ^ { \\prime } } \\\\ { \\displaystyle \\qquad = s \\int _ { \\mathbb R } f ( s ^ { - 1 } x ^ { \\prime } ) \\psi ( s ( s ^ { - 1 } x ^ { \\prime } - s ^ { - 1 } x ) ) d ( s ^ { - 1 } x ^ { \\prime } ) = s L _ { s } [ f \\star L _ { s ^ { - 1 } } [ \\psi ] ] ( x ) } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "cecee878bf94f360d23a829c220e0eba59c655aa6e7d443c42052b97c09c1b7c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 30,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 131,
+ 512,
+ 480,
+ 530.3333333333334
+ ],
+ "spans": [],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 131,
+ 530.3333333333334,
+ 480,
+ 548.6666666666667
+ ],
+ "spans": [],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 131,
+ 548.6666666666667,
+ 480,
+ 567.0000000000001
+ ],
+ "spans": [],
+ "index": 31
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 576,
+ 504,
+ 609
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 574,
+ 506,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 574,
+ 506,
+ 589
+ ],
+ "score": 1.0,
+ "content": "In other words, convolution of the downscaled function with a filter can be expressed through a",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 599
+ ],
+ "score": 1.0,
+ "content": "convolution of the function with the correspondingly upscaled filter where downscaling is performed",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 107,
+ 598,
+ 453,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 598,
+ 453,
+ 610
+ ],
+ "score": 1.0,
+ "content": "afterwards. Equation 2 shows us that the standard convolution is not scale-equivariant.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 615,
+ 504,
+ 637
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 408,
+ 628
+ ],
+ "score": 1.0,
+ "content": "Steerable Filters In order to make computations simpler, we reparametrize",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 408,
+ 614,
+ 501,
+ 627
+ ],
+ "score": 0.93,
+ "content": "\\psi _ { \\sigma } ( x ) = \\sigma ^ { - 1 } \\psi ( \\sigma ^ { - 1 } x )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 613,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 624,
+ 244,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 624,
+ 244,
+ 639
+ ],
+ "score": 1.0,
+ "content": "which has the following property:",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 223,
+ 642,
+ 388,
+ 657
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 223,
+ 642,
+ 388,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 223,
+ 642,
+ 388,
+ 657
+ ],
+ "score": 0.91,
+ "content": "L _ { s ^ { - 1 } } [ \\psi _ { \\sigma } ] ( x ) = \\psi _ { \\sigma } ( s x ) = s ^ { - 1 } \\psi _ { s ^ { - 1 } \\sigma } ( x )",
+ "type": "interline_equation",
+ "image_path": "5c5e01cd9c1c7f93935e09fa08fbca4b21f2258a0937e7a59b257eb0bc5c51cc.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 37,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 223,
+ 642,
+ 388,
+ 657
+ ],
+ "spans": [],
+ "index": 37
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 667,
+ 265,
+ 680
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 266,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 266,
+ 681
+ ],
+ "score": 1.0,
+ "content": "It gives a shorter version of Equation 2:",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 38
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 247,
+ 685,
+ 364,
+ 698
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 247,
+ 685,
+ 364,
+ 698
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 247,
+ 685,
+ 364,
+ 698
+ ],
+ "score": 0.91,
+ "content": "L _ { s } [ f ] \\star \\psi _ { \\sigma } = L _ { s } [ f \\star \\psi _ { s ^ { - 1 } \\sigma } ]",
+ "type": "interline_equation",
+ "image_path": "7651f206035b5725275d06c6bc6580fff017e6d183066a68fe1bc3c8467c7de9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 39,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 247,
+ 685,
+ 364,
+ 698
+ ],
+ "spans": [],
+ "index": 39
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 709,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "We will refer to such a parameterization of filters as Steerable Filters because the scaling of these",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 505,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 505,
+ 734
+ ],
+ "score": 1.0,
+ "content": "filters is the transformation of its parameters. Note that we may construct steerable filters from any",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40.5
+ }
+ ],
+ "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 2020",
+ "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": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 149
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": "In this paper we develop a theory of scale-equivariant networks. We demonstrate the concept of",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "steerable filter parametrization which allows for scaling without the need for tensor resizing. Then",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 105,
+ 504,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 105,
+ 504,
+ 117
+ ],
+ "score": 1.0,
+ "content": "we derive scale-equivariant convolution and demonstrate a fast algorithm for its implementation.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 505,
+ 128
+ ],
+ "score": 1.0,
+ "content": "Furthermore, we experiment to determine to what degree the mathematical properties actually hold",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 138
+ ],
+ "score": 1.0,
+ "content": "true. Finally, we conduct a set of experiments comparing our model with other methods for scale",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 138,
+ 267,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 138,
+ 267,
+ 149
+ ],
+ "score": 1.0,
+ "content": "equivariance and local scale invariance.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 2.5,
+ "bbox_fs": [
+ 105,
+ 82,
+ 506,
+ 149
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 168,
+ 487,
+ 180
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 168,
+ 488,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 168,
+ 488,
+ 182
+ ],
+ "score": 1.0,
+ "content": "The proposed model has the following advantages compared to other scale-equivariant models:",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 106,
+ 168,
+ 488,
+ 182
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 129,
+ 189,
+ 505,
+ 275
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 190,
+ 505,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 190,
+ 505,
+ 201
+ ],
+ "score": 1.0,
+ "content": "1. It is equivariant to scale transformations with arbitrary discrete scale factors and is not",
+ "type": "text"
+ }
+ ],
+ "index": 7,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 141,
+ 200,
+ 431,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 200,
+ 431,
+ 213
+ ],
+ "score": 1.0,
+ "content": "limited to either integer scales or scales tailored by the image pixel grid.",
+ "type": "text"
+ }
+ ],
+ "index": 8,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 129,
+ 215,
+ 505,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 215,
+ 505,
+ 228
+ ],
+ "score": 1.0,
+ "content": "2. It does not rely on any image resampling techniques during training, and therefore, pro-",
+ "type": "text"
+ }
+ ],
+ "index": 9,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 141,
+ 226,
+ 457,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 226,
+ 457,
+ 239
+ ],
+ "score": 1.0,
+ "content": "duces deep scale-equivariant representations free of any interpolation artifacts.",
+ "type": "text"
+ }
+ ],
+ "index": 10,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 129,
+ 241,
+ 505,
+ 254
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 241,
+ 505,
+ 254
+ ],
+ "score": 1.0,
+ "content": "3. The algorithm is based on the combination of tensor expansion and 2-dimensional convo-",
+ "type": "text"
+ }
+ ],
+ "index": 11,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 141,
+ 252,
+ 506,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 252,
+ 506,
+ 265
+ ],
+ "score": 1.0,
+ "content": "lution, and demonstrates the same computation time as the general CNN with a comparable",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 142,
+ 263,
+ 188,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 263,
+ 188,
+ 275
+ ],
+ "score": 1.0,
+ "content": "filter bank.",
+ "type": "text"
+ }
+ ],
+ "index": 13,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 10,
+ "bbox_fs": [
+ 129,
+ 190,
+ 506,
+ 275
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 290,
+ 208,
+ 303
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 289,
+ 209,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 289,
+ 209,
+ 306
+ ],
+ "score": 1.0,
+ "content": "2 PRELIMINARIES",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 315,
+ 505,
+ 360
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 505,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 505,
+ 329
+ ],
+ "score": 1.0,
+ "content": "Before we move into scale-equivariant mappings, we discuss some aspects of equivariance, scaling",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 327,
+ 504,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 327,
+ 504,
+ 339
+ ],
+ "score": 1.0,
+ "content": "transformations, symmetry groups, and the functions defined on them. For simplicity, in this sec-",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 338,
+ 505,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 338,
+ 505,
+ 350
+ ],
+ "score": 1.0,
+ "content": "tion, we consider only 1-dimensional functions. The generalization to higher-dimensional cases is",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 348,
+ 173,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 348,
+ 173,
+ 361
+ ],
+ "score": 1.0,
+ "content": "straightforward.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 16.5,
+ "bbox_fs": [
+ 105,
+ 315,
+ 505,
+ 361
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 365,
+ 504,
+ 388
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 365,
+ 505,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 365,
+ 289,
+ 378
+ ],
+ "score": 1.0,
+ "content": "Equivariance Let us consider some mapping",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 290,
+ 368,
+ 296,
+ 377
+ ],
+ "score": 0.74,
+ "content": "g",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 297,
+ 365,
+ 390,
+ 378
+ ],
+ "score": 1.0,
+ "content": ". It is equivariant under",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 390,
+ 366,
+ 402,
+ 376
+ ],
+ "score": 0.88,
+ "content": "L _ { \\theta }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 402,
+ 365,
+ 505,
+ 378
+ ],
+ "score": 1.0,
+ "content": "if and only if there exists",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 107,
+ 375,
+ 477,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 377,
+ 119,
+ 389
+ ],
+ "score": 0.89,
+ "content": "L _ { \\theta } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 119,
+ 375,
+ 158,
+ 390
+ ],
+ "score": 1.0,
+ "content": "such that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 159,
+ 376,
+ 225,
+ 389
+ ],
+ "score": 0.93,
+ "content": "g \\circ L _ { \\theta } = L _ { \\theta } ^ { \\prime } \\circ g",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 225,
+ 375,
+ 259,
+ 390
+ ],
+ "score": 1.0,
+ "content": ". In case",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 376,
+ 272,
+ 389
+ ],
+ "score": 0.92,
+ "content": "L _ { \\theta } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 272,
+ 375,
+ 420,
+ 390
+ ],
+ "score": 1.0,
+ "content": "is the identity mapping, the function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 420,
+ 378,
+ 426,
+ 388
+ ],
+ "score": 0.79,
+ "content": "g",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 427,
+ 375,
+ 477,
+ 390
+ ],
+ "score": 1.0,
+ "content": "is invariant.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 19.5,
+ "bbox_fs": [
+ 106,
+ 365,
+ 505,
+ 390
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 393,
+ 505,
+ 426
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 393,
+ 505,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 393,
+ 505,
+ 405
+ ],
+ "score": 1.0,
+ "content": "In this paper we consider scaling transformations. In order to guarantee the equivariance of the",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 416
+ ],
+ "score": 1.0,
+ "content": "predictions to such transformations, and to improve the performance of the model, we seek to incor-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 416,
+ 276,
+ 427
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 416,
+ 276,
+ 427
+ ],
+ "score": 1.0,
+ "content": "porate this property directly inside CNNs.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22,
+ "bbox_fs": [
+ 105,
+ 393,
+ 505,
+ 427
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 432,
+ 435,
+ 444
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 431,
+ 435,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 431,
+ 209,
+ 445
+ ],
+ "score": 1.0,
+ "content": "Scaling Given a function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 209,
+ 432,
+ 254,
+ 444
+ ],
+ "score": 0.92,
+ "content": "f : \\mathbb { R } \\to \\mathbb { R }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 255,
+ 431,
+ 435,
+ 445
+ ],
+ "score": 1.0,
+ "content": ", a scale transformation is defined as follows:",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 24,
+ "bbox_fs": [
+ 105,
+ 431,
+ 435,
+ 445
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 240,
+ 448,
+ 371,
+ 463
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 240,
+ 448,
+ 371,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 240,
+ 448,
+ 371,
+ 463
+ ],
+ "score": 0.91,
+ "content": "L _ { s } [ f ] ( x ) = f ( s ^ { - 1 } x ) , \\quad \\forall s > 0",
+ "type": "interline_equation",
+ "image_path": "13908d2afa72d06c5f750fd9545ff5fd44438e6667a9eef4ad001702836f5831.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 25,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 240,
+ 448,
+ 371,
+ 463
+ ],
+ "spans": [],
+ "index": 25
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 474,
+ 505,
+ 507
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 474,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 474,
+ 199,
+ 486
+ ],
+ "score": 1.0,
+ "content": "We refer to cases with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 199,
+ 475,
+ 224,
+ 484
+ ],
+ "score": 0.9,
+ "content": "s > 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 224,
+ 474,
+ 341,
+ 486
+ ],
+ "score": 1.0,
+ "content": "as upscale and to cases with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 342,
+ 475,
+ 367,
+ 484
+ ],
+ "score": 0.9,
+ "content": "s < 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 367,
+ 474,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "as downscale. If we convolve the",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 485,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 485,
+ 284,
+ 497
+ ],
+ "score": 1.0,
+ "content": "downscaled function with an arbitrary filter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 284,
+ 486,
+ 292,
+ 497
+ ],
+ "score": 0.85,
+ "content": "\\psi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 293,
+ 485,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "and perform a simple change of variables inside the",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 495,
+ 265,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 495,
+ 265,
+ 509
+ ],
+ "score": 1.0,
+ "content": "integral, we get the following property:",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27,
+ "bbox_fs": [
+ 105,
+ 474,
+ 505,
+ 509
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 131,
+ 512,
+ 480,
+ 567
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 512,
+ 480,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 512,
+ 480,
+ 567
+ ],
+ "score": 0.95,
+ "content": "\\begin{array} { l } { \\displaystyle [ L _ { s } [ f ] \\star \\psi ] ( x ) = \\int _ { \\mathbb R } L _ { s } [ f ] ( x ^ { \\prime } ) \\psi ( x ^ { \\prime } - x ) d x ^ { \\prime } = \\int _ { \\mathbb R } f ( s ^ { - 1 } x ^ { \\prime } ) \\psi ( x ^ { \\prime } - x ) d x ^ { \\prime } } \\\\ { \\displaystyle \\qquad = s \\int _ { \\mathbb R } f ( s ^ { - 1 } x ^ { \\prime } ) \\psi ( s ( s ^ { - 1 } x ^ { \\prime } - s ^ { - 1 } x ) ) d ( s ^ { - 1 } x ^ { \\prime } ) = s L _ { s } [ f \\star L _ { s ^ { - 1 } } [ \\psi ] ] ( x ) } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "cecee878bf94f360d23a829c220e0eba59c655aa6e7d443c42052b97c09c1b7c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 30,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 131,
+ 512,
+ 480,
+ 530.3333333333334
+ ],
+ "spans": [],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 131,
+ 530.3333333333334,
+ 480,
+ 548.6666666666667
+ ],
+ "spans": [],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 131,
+ 548.6666666666667,
+ 480,
+ 567.0000000000001
+ ],
+ "spans": [],
+ "index": 31
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 576,
+ 504,
+ 609
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 574,
+ 506,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 574,
+ 506,
+ 589
+ ],
+ "score": 1.0,
+ "content": "In other words, convolution of the downscaled function with a filter can be expressed through a",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 599
+ ],
+ "score": 1.0,
+ "content": "convolution of the function with the correspondingly upscaled filter where downscaling is performed",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 107,
+ 598,
+ 453,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 598,
+ 453,
+ 610
+ ],
+ "score": 1.0,
+ "content": "afterwards. Equation 2 shows us that the standard convolution is not scale-equivariant.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 33,
+ "bbox_fs": [
+ 105,
+ 574,
+ 506,
+ 610
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 615,
+ 504,
+ 637
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 408,
+ 628
+ ],
+ "score": 1.0,
+ "content": "Steerable Filters In order to make computations simpler, we reparametrize",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 408,
+ 614,
+ 501,
+ 627
+ ],
+ "score": 0.93,
+ "content": "\\psi _ { \\sigma } ( x ) = \\sigma ^ { - 1 } \\psi ( \\sigma ^ { - 1 } x )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 613,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 624,
+ 244,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 624,
+ 244,
+ 639
+ ],
+ "score": 1.0,
+ "content": "which has the following property:",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35.5,
+ "bbox_fs": [
+ 105,
+ 613,
+ 505,
+ 639
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 223,
+ 642,
+ 388,
+ 657
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 223,
+ 642,
+ 388,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 223,
+ 642,
+ 388,
+ 657
+ ],
+ "score": 0.91,
+ "content": "L _ { s ^ { - 1 } } [ \\psi _ { \\sigma } ] ( x ) = \\psi _ { \\sigma } ( s x ) = s ^ { - 1 } \\psi _ { s ^ { - 1 } \\sigma } ( x )",
+ "type": "interline_equation",
+ "image_path": "5c5e01cd9c1c7f93935e09fa08fbca4b21f2258a0937e7a59b257eb0bc5c51cc.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 37,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 223,
+ 642,
+ 388,
+ 657
+ ],
+ "spans": [],
+ "index": 37
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 667,
+ 265,
+ 680
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 266,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 266,
+ 681
+ ],
+ "score": 1.0,
+ "content": "It gives a shorter version of Equation 2:",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 38,
+ "bbox_fs": [
+ 106,
+ 667,
+ 266,
+ 681
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 247,
+ 685,
+ 364,
+ 698
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 247,
+ 685,
+ 364,
+ 698
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 247,
+ 685,
+ 364,
+ 698
+ ],
+ "score": 0.91,
+ "content": "L _ { s } [ f ] \\star \\psi _ { \\sigma } = L _ { s } [ f \\star \\psi _ { s ^ { - 1 } \\sigma } ]",
+ "type": "interline_equation",
+ "image_path": "7651f206035b5725275d06c6bc6580fff017e6d183066a68fe1bc3c8467c7de9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 39,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 247,
+ 685,
+ 364,
+ 698
+ ],
+ "spans": [],
+ "index": 39
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 709,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "We will refer to such a parameterization of filters as Steerable Filters because the scaling of these",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 505,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 505,
+ 734
+ ],
+ "score": 1.0,
+ "content": "filters is the transformation of its parameters. Note that we may construct steerable filters from any",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 94
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 94
+ ],
+ "score": 1.0,
+ "content": "function. This has the important consequence that it does not restrict our approach. Rather it will",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "make the analysis easier for discrete data. Moreover, note that any linear combination of steerable",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 105,
+ 199,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 105,
+ 199,
+ 116
+ ],
+ "score": 1.0,
+ "content": "filters is still steerable.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 105,
+ 709,
+ 505,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 82,
+ 504,
+ 115
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 94
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 94
+ ],
+ "score": 1.0,
+ "content": "function. This has the important consequence that it does not restrict our approach. Rather it will",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "make the analysis easier for discrete data. Moreover, note that any linear combination of steerable",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 105,
+ 199,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 105,
+ 199,
+ 116
+ ],
+ "score": 1.0,
+ "content": "filters is still steerable.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 121,
+ 505,
+ 222
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 121,
+ 506,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 121,
+ 397,
+ 135
+ ],
+ "score": 1.0,
+ "content": "Scale-Translation Group All possible scales form the scaling group",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 397,
+ 122,
+ 405,
+ 131
+ ],
+ "score": 0.78,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 405,
+ 121,
+ 506,
+ 135
+ ],
+ "score": 1.0,
+ "content": ". Here we consider the",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 132,
+ 506,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 132,
+ 282,
+ 145
+ ],
+ "score": 1.0,
+ "content": "discrete scale group, i.e. scales of the form",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 282,
+ 132,
+ 381,
+ 144
+ ],
+ "score": 0.91,
+ "content": "\\dots a ^ { - 1 } , a ^ { - 1 } , 1 , a , a ^ { \\tilde { 2 } } , \\tilde { \\dots } .",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 381,
+ 132,
+ 425,
+ 145
+ ],
+ "score": 1.0,
+ "content": "with base",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 426,
+ 135,
+ 432,
+ 142
+ ],
+ "score": 0.69,
+ "content": "a",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 433,
+ 132,
+ 506,
+ 145
+ ],
+ "score": 1.0,
+ "content": "as a parameter of",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 143,
+ 506,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 143,
+ 506,
+ 156
+ ],
+ "score": 1.0,
+ "content": "our method. Analysis of this group by itself breaks the translation equivariance of CNNs. Thus we",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 154,
+ 505,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 154,
+ 505,
+ 167
+ ],
+ "score": 1.0,
+ "content": "seek to incorporate scale and translation symmetries into CNNs, and, therefore consider the Scale-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 164,
+ 505,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 164,
+ 182,
+ 178
+ ],
+ "score": 1.0,
+ "content": "Translation Group",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 182,
+ 166,
+ 192,
+ 175
+ ],
+ "score": 0.72,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 192,
+ 164,
+ 379,
+ 178
+ ],
+ "score": 1.0,
+ "content": ". It is a semidirect product of the scaling group",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 380,
+ 166,
+ 388,
+ 175
+ ],
+ "score": 0.8,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 388,
+ 164,
+ 505,
+ 178
+ ],
+ "score": 1.0,
+ "content": "and the group of translations",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 176,
+ 505,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 176,
+ 136,
+ 186
+ ],
+ "score": 0.88,
+ "content": "T \\cong \\mathbb { R }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 136,
+ 176,
+ 205,
+ 189
+ ],
+ "score": 1.0,
+ "content": ". In other words:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 176,
+ 313,
+ 188
+ ],
+ "score": 0.91,
+ "content": "H = \\{ ( s , t ) | s \\in S , t \\in T \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 313,
+ 176,
+ 505,
+ 189
+ ],
+ "score": 1.0,
+ "content": ". For multiplication of group elements, we have",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 107,
+ 187,
+ 505,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 188,
+ 250,
+ 200
+ ],
+ "score": 0.9,
+ "content": "( s _ { 2 } , t _ { 2 } ) \\cdot ( s _ { 1 } , t _ { 1 } ) = ( s _ { 2 } s _ { 1 } , s _ { 2 } t _ { 1 } + t _ { 2 } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 251,
+ 187,
+ 327,
+ 202
+ ],
+ "score": 1.0,
+ "content": "and for the inverse",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 327,
+ 187,
+ 500,
+ 201
+ ],
+ "score": 0.91,
+ "content": "( s _ { 2 } , t _ { 2 } ) ^ { - 1 } \\cdot ( s _ { 1 } , t _ { 1 } ) = ( s _ { 2 } ^ { - 1 } s _ { 1 } , s _ { 2 } ^ { - 1 } ( t _ { 1 } - t _ { 2 } ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 187,
+ 505,
+ 202
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 104,
+ 198,
+ 505,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 198,
+ 443,
+ 213
+ ],
+ "score": 1.0,
+ "content": "Additionally, for the corresponding scaling and translation transformations, we have",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 443,
+ 200,
+ 505,
+ 211
+ ],
+ "score": 0.88,
+ "content": "L _ { s t } = L _ { s } L _ { t } \\neq",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 210,
+ 347,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 210,
+ 129,
+ 222
+ ],
+ "score": 0.9,
+ "content": "L _ { t } L _ { s }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 129,
+ 210,
+ 347,
+ 223
+ ],
+ "score": 1.0,
+ "content": ", which means that the order of the operations matters.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 227,
+ 505,
+ 284
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 227,
+ 505,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 227,
+ 421,
+ 239
+ ],
+ "score": 1.0,
+ "content": "From now on, we will work with functions defined on groups, i.e. mappings",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 421,
+ 227,
+ 456,
+ 237
+ ],
+ "score": 0.9,
+ "content": "H \\to \\mathbb { R }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 457,
+ 227,
+ 505,
+ 239
+ ],
+ "score": 1.0,
+ "content": ". Note, that",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 238,
+ 505,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 238,
+ 172,
+ 250
+ ],
+ "score": 1.0,
+ "content": "simple function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 173,
+ 239,
+ 222,
+ 250
+ ],
+ "score": 0.93,
+ "content": "f : \\mathbb { R } \\mathbb { R }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 223,
+ 238,
+ 371,
+ 250
+ ],
+ "score": 1.0,
+ "content": "may be considered as a function on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 372,
+ 239,
+ 382,
+ 248
+ ],
+ "score": 0.8,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 382,
+ 238,
+ 505,
+ 250
+ ],
+ "score": 1.0,
+ "content": "with constant value along the",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 107,
+ 248,
+ 505,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 250,
+ 114,
+ 259
+ ],
+ "score": 0.81,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 115,
+ 248,
+ 340,
+ 262
+ ],
+ "score": 1.0,
+ "content": "axis. Therefore, Equation 4 holds true for functions on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 340,
+ 250,
+ 350,
+ 259
+ ],
+ "score": 0.83,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 351,
+ 248,
+ 505,
+ 262
+ ],
+ "score": 1.0,
+ "content": "as well. One thing we should keep in",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 260,
+ 505,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 260,
+ 222,
+ 272
+ ],
+ "score": 1.0,
+ "content": "mind is that when we apply",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 222,
+ 261,
+ 234,
+ 271
+ ],
+ "score": 0.88,
+ "content": "L _ { s }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 235,
+ 260,
+ 300,
+ 272
+ ],
+ "score": 1.0,
+ "content": "to functions on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 300,
+ 261,
+ 311,
+ 270
+ ],
+ "score": 0.83,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 311,
+ 260,
+ 330,
+ 272
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 330,
+ 261,
+ 338,
+ 270
+ ],
+ "score": 0.83,
+ "content": "\\mathbb { R }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 339,
+ 260,
+ 505,
+ 272
+ ],
+ "score": 1.0,
+ "content": "we use different notations. For example",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 270,
+ 427,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 197,
+ 284
+ ],
+ "score": 0.92,
+ "content": "L _ { s } [ f ] ( x ^ { \\prime } ) = f ( s ^ { - 1 } x ^ { \\prime } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 198,
+ 270,
+ 215,
+ 284
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 216,
+ 271,
+ 427,
+ 284
+ ],
+ "score": 0.83,
+ "content": "\\begin{array} { r } { \\bar { L } _ { s } [ f ] ( s ^ { \\prime } , t ^ { \\prime } ) = f ( ( s , 0 ) ^ { - 1 } ( s ^ { \\prime } , t ^ { \\prime } ) ) = f ( s ^ { - 1 } s ^ { \\prime } , s ^ { - 1 } t ^ { \\prime } ) } \\end{array}",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 287,
+ 505,
+ 311
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 287,
+ 505,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 287,
+ 304,
+ 301
+ ],
+ "score": 1.0,
+ "content": "Group-Equivariant Convolution Given group",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 288,
+ 313,
+ 298
+ ],
+ "score": 0.82,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 313,
+ 287,
+ 393,
+ 301
+ ],
+ "score": 1.0,
+ "content": "and two functions",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 393,
+ 289,
+ 401,
+ 300
+ ],
+ "score": 0.86,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 401,
+ 287,
+ 421,
+ 301
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 421,
+ 289,
+ 430,
+ 299
+ ],
+ "score": 0.85,
+ "content": "\\psi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 430,
+ 287,
+ 491,
+ 301
+ ],
+ "score": 1.0,
+ "content": "defined on it,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 492,
+ 288,
+ 501,
+ 298
+ ],
+ "score": 0.77,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 287,
+ 505,
+ 301
+ ],
+ "score": 1.0,
+ "content": "-",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 299,
+ 248,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 299,
+ 248,
+ 312
+ ],
+ "score": 1.0,
+ "content": "equivariant convolution is given by",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 17.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 165,
+ 313,
+ 445,
+ 340
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 165,
+ 313,
+ 445,
+ 340
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 165,
+ 313,
+ 445,
+ 340
+ ],
+ "score": 0.94,
+ "content": "[ f \\star _ { G } \\psi ] ( g ) = \\int _ { G } f ( g ^ { \\prime } ) L _ { g } [ \\psi ] ( g ^ { \\prime } ) d \\mu ( g ^ { \\prime } ) = \\int _ { G } f ( g ^ { \\prime } ) \\psi ( g ^ { - 1 } g ^ { \\prime } ) d \\mu ( g ^ { \\prime } )",
+ "type": "interline_equation",
+ "image_path": "5988433815b304c1229351695466f4b8fe3fc481fadcd0844c1eef82a7d8bd97.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 19,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 165,
+ 313,
+ 445,
+ 340
+ ],
+ "spans": [],
+ "index": 19
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 344,
+ 505,
+ 389
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 506,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 129,
+ 357
+ ],
+ "score": 1.0,
+ "content": "Here",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 129,
+ 344,
+ 151,
+ 356
+ ],
+ "score": 0.91,
+ "content": "\\mu ( g ^ { \\prime } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 152,
+ 343,
+ 457,
+ 357
+ ],
+ "score": 1.0,
+ "content": "is the Haar measure also known as invariant measure Folland (2016). For",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 457,
+ 344,
+ 489,
+ 354
+ ],
+ "score": 0.91,
+ "content": "T \\cong \\mathbb { R }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 489,
+ 343,
+ 506,
+ 357
+ ],
+ "score": 1.0,
+ "content": "we",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 354,
+ 506,
+ 368
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 354,
+ 127,
+ 368
+ ],
+ "score": 1.0,
+ "content": "have",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 128,
+ 355,
+ 182,
+ 367
+ ],
+ "score": 0.92,
+ "content": "d \\mu ( g ^ { \\prime } ) = d g ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 183,
+ 354,
+ 506,
+ 368
+ ],
+ "score": 1.0,
+ "content": ". For discrete groups, the Haar measure is the counting measure, and integration",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 366,
+ 504,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 497,
+ 379
+ ],
+ "score": 1.0,
+ "content": "becomes a discrete sum. This formula tells us that the output of the convolution evaluated at point",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 498,
+ 370,
+ 504,
+ 378
+ ],
+ "score": 0.73,
+ "content": "g",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 104,
+ 376,
+ 413,
+ 391
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 376,
+ 272,
+ 391
+ ],
+ "score": 1.0,
+ "content": "is the inner product between the function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 272,
+ 378,
+ 279,
+ 389
+ ],
+ "score": 0.86,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 280,
+ 376,
+ 384,
+ 391
+ ],
+ "score": 1.0,
+ "content": "and the transformed filter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 384,
+ 377,
+ 409,
+ 390
+ ],
+ "score": 0.92,
+ "content": "L _ { g } [ \\psi ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 409,
+ 376,
+ 413,
+ 391
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 21.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 404,
+ 296,
+ 417
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 402,
+ 298,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 402,
+ 298,
+ 420
+ ],
+ "score": 1.0,
+ "content": "3 SCALE-EQUIVARIANT MAPPINGS",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 428,
+ 384,
+ 441
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 385,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 385,
+ 443
+ ],
+ "score": 1.0,
+ "content": "Now we define the main building blocks of scale-equivariant models.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 446,
+ 505,
+ 491
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 458
+ ],
+ "score": 1.0,
+ "content": "Scale Convolution In order to derive scale convolution, we start from group equivariant convolution",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 457,
+ 505,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 457,
+ 127,
+ 470
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 127,
+ 457,
+ 161,
+ 467
+ ],
+ "score": 0.89,
+ "content": "G = H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 162,
+ 457,
+ 505,
+ 470
+ ],
+ "score": 1.0,
+ "content": ". We first use the property of semidirect product of groups which splits the integral,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 467,
+ 506,
+ 481
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 467,
+ 506,
+ 481
+ ],
+ "score": 1.0,
+ "content": "then choose the appropriate Haar measures, and finally use the properties of steerable filters. Given",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 478,
+ 489,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 157,
+ 492
+ ],
+ "score": 1.0,
+ "content": "the function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 157,
+ 479,
+ 184,
+ 491
+ ],
+ "score": 0.92,
+ "content": "f ( s , t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 185,
+ 478,
+ 269,
+ 492
+ ],
+ "score": 1.0,
+ "content": "and a steerable filter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 269,
+ 479,
+ 302,
+ 491
+ ],
+ "score": 0.93,
+ "content": "\\psi _ { \\sigma } ( s , t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 302,
+ 478,
+ 348,
+ 492
+ ],
+ "score": 1.0,
+ "content": "defined on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 348,
+ 479,
+ 358,
+ 489
+ ],
+ "score": 0.82,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 358,
+ 478,
+ 489,
+ 492
+ ],
+ "score": 1.0,
+ "content": ", a scale convolution is given by:",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 27.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 119,
+ 495,
+ 479,
+ 551
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 495,
+ 479,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 495,
+ 479,
+ 551
+ ],
+ "score": 0.93,
+ "content": "\\begin{array} { l } { { [ f \\star _ { \\cal H } \\psi _ { \\sigma } ] ( s , t ) = \\displaystyle \\int _ { \\cal S } \\int _ { \\cal T } f ( s ^ { \\prime } , t ^ { \\prime } ) { \\cal L } _ { s t } [ \\psi _ { \\sigma } ] ( s ^ { \\prime } , t ^ { \\prime } ) d \\mu ( s ^ { \\prime } ) d \\mu ( t ^ { \\prime } ) } } \\\\ { { { } } } \\\\ { { { } = \\displaystyle \\sum _ { s ^ { \\prime } } \\displaystyle \\int _ { \\cal T } f ( s ^ { \\prime } , t ^ { \\prime } ) \\psi _ { s \\sigma } ( s ^ { - 1 } s ^ { \\prime } , t ^ { \\prime } - t ) d t ^ { \\prime } = \\displaystyle \\sum _ { s ^ { \\prime } } [ f ( s ^ { \\prime } , \\cdot ) \\star \\psi _ { s \\sigma } ( s ^ { - 1 } s ^ { \\prime } , \\cdot ) ] ( t ) } } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "8cb3ab9dc952636da09b958c251b9f4946d4bb834f2be16247217b30e4ccb17c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 31,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 119,
+ 495,
+ 479,
+ 513.6666666666666
+ ],
+ "spans": [],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 119,
+ 513.6666666666666,
+ 479,
+ 532.3333333333333
+ ],
+ "spans": [],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 119,
+ 532.3333333333333,
+ 479,
+ 550.9999999999999
+ ],
+ "spans": [],
+ "index": 32
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 560,
+ 362,
+ 572
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 559,
+ 363,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 559,
+ 185,
+ 573
+ ],
+ "score": 1.0,
+ "content": "And for the case of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 186,
+ 561,
+ 200,
+ 571
+ ],
+ "score": 0.89,
+ "content": "C _ { \\mathrm { i n } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 200,
+ 559,
+ 241,
+ 573
+ ],
+ "score": 1.0,
+ "content": "input and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 241,
+ 561,
+ 259,
+ 571
+ ],
+ "score": 0.9,
+ "content": "C _ { \\mathrm { o u t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 259,
+ 559,
+ 363,
+ 573
+ ],
+ "score": 1.0,
+ "content": "output channels we have:",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 146,
+ 576,
+ 465,
+ 610
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 146,
+ 576,
+ 465,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 576,
+ 465,
+ 610
+ ],
+ "score": 0.94,
+ "content": "[ f \\star _ { H } \\psi _ { \\sigma } ] _ { m } ( s , t ) = \\sum _ { n = 1 } ^ { C _ { \\mathrm { i n } } } \\sum _ { s ^ { \\prime } } [ f _ { n } ( s ^ { \\prime } , \\cdot ) \\star \\psi _ { n , m , s \\sigma } ( s ^ { - 1 } s ^ { \\prime } , \\cdot ) ] ( t ) , \\quad m = 1 \\ldots C _ { \\mathrm { o u t } }",
+ "type": "interline_equation",
+ "image_path": "c5aa5fbbeeac0f26cf646a9653a08ff81f5b335b69d41fab9884696d6ccab114.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 35,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 146,
+ 576,
+ 465,
+ 587.3333333333334
+ ],
+ "spans": [],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 146,
+ 587.3333333333334,
+ 465,
+ 598.6666666666667
+ ],
+ "spans": [],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 146,
+ 598.6666666666667,
+ 465,
+ 610.0000000000001
+ ],
+ "spans": [],
+ "index": 36
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 613,
+ 503,
+ 625
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 612,
+ 504,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 612,
+ 398,
+ 626
+ ],
+ "score": 1.0,
+ "content": "The proof of the equivariance of this convolution to transformations from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 398,
+ 614,
+ 408,
+ 623
+ ],
+ "score": 0.82,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 408,
+ 612,
+ 504,
+ 626
+ ],
+ "score": 1.0,
+ "content": "is given in Appendix A.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 630,
+ 505,
+ 686
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 630,
+ 505,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 630,
+ 505,
+ 642
+ ],
+ "score": 1.0,
+ "content": "Kondor & Trivedi (2018) prove that a feed-forward neural network is equivariant to transformations",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 641,
+ 505,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 641,
+ 129,
+ 653
+ ],
+ "score": 1.0,
+ "content": "from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 129,
+ 642,
+ 138,
+ 651
+ ],
+ "score": 0.81,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 139,
+ 641,
+ 288,
+ 653
+ ],
+ "score": 1.0,
+ "content": "if and only if it is constructed from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 288,
+ 642,
+ 297,
+ 651
+ ],
+ "score": 0.27,
+ "content": "\\mathbf { G }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 297,
+ 641,
+ 505,
+ 653
+ ],
+ "score": 1.0,
+ "content": "-equivariant convolutional layers. Thus Equation 7",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 653,
+ 505,
+ 665
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 653,
+ 505,
+ 665
+ ],
+ "score": 1.0,
+ "content": "shows the most general form of scale-equivariant layers which allows for building scale-equivariant",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 662,
+ 506,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 662,
+ 280,
+ 677
+ ],
+ "score": 1.0,
+ "content": "convolutional networks with such choice of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 280,
+ 664,
+ 288,
+ 673
+ ],
+ "score": 0.75,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 288,
+ 662,
+ 506,
+ 677
+ ],
+ "score": 1.0,
+ "content": ". We will refer to models using scale-equivariant layers",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 673,
+ 426,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 673,
+ 398,
+ 687
+ ],
+ "score": 1.0,
+ "content": "with steerable filters as Scale-Equivariant Steerable Networks, or shortly",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 398,
+ 674,
+ 426,
+ 685
+ ],
+ "score": 0.33,
+ "content": "S \\bar { E } S N ^ { 1 }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 40
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 691,
+ 504,
+ 714
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 690,
+ 506,
+ 704
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 690,
+ 506,
+ 704
+ ],
+ "score": 1.0,
+ "content": "Nonlinearities In order to guarantee the equivariance of the network to scale transformations, we",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 702,
+ 506,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 702,
+ 506,
+ 714
+ ],
+ "score": 1.0,
+ "content": "use scale equivariant nonlinearities. We are free to use simple point-wise nonlinearities. Indeed,",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43.5
+ }
+ ],
+ "page_idx": 2,
+ "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 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 120,
+ 722,
+ 197,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 120,
+ 720,
+ 199,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 120,
+ 720,
+ 199,
+ 733
+ ],
+ "score": 1.0,
+ "content": "1pronounced ‘season’",
+ "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": "text",
+ "bbox": [
+ 108,
+ 82,
+ 504,
+ 115
+ ],
+ "lines": [],
+ "index": 1,
+ "bbox_fs": [
+ 106,
+ 82,
+ 505,
+ 116
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 121,
+ 505,
+ 222
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 121,
+ 506,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 121,
+ 397,
+ 135
+ ],
+ "score": 1.0,
+ "content": "Scale-Translation Group All possible scales form the scaling group",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 397,
+ 122,
+ 405,
+ 131
+ ],
+ "score": 0.78,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 405,
+ 121,
+ 506,
+ 135
+ ],
+ "score": 1.0,
+ "content": ". Here we consider the",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 132,
+ 506,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 132,
+ 282,
+ 145
+ ],
+ "score": 1.0,
+ "content": "discrete scale group, i.e. scales of the form",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 282,
+ 132,
+ 381,
+ 144
+ ],
+ "score": 0.91,
+ "content": "\\dots a ^ { - 1 } , a ^ { - 1 } , 1 , a , a ^ { \\tilde { 2 } } , \\tilde { \\dots } .",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 381,
+ 132,
+ 425,
+ 145
+ ],
+ "score": 1.0,
+ "content": "with base",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 426,
+ 135,
+ 432,
+ 142
+ ],
+ "score": 0.69,
+ "content": "a",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 433,
+ 132,
+ 506,
+ 145
+ ],
+ "score": 1.0,
+ "content": "as a parameter of",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 143,
+ 506,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 143,
+ 506,
+ 156
+ ],
+ "score": 1.0,
+ "content": "our method. Analysis of this group by itself breaks the translation equivariance of CNNs. Thus we",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 154,
+ 505,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 154,
+ 505,
+ 167
+ ],
+ "score": 1.0,
+ "content": "seek to incorporate scale and translation symmetries into CNNs, and, therefore consider the Scale-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 164,
+ 505,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 164,
+ 182,
+ 178
+ ],
+ "score": 1.0,
+ "content": "Translation Group",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 182,
+ 166,
+ 192,
+ 175
+ ],
+ "score": 0.72,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 192,
+ 164,
+ 379,
+ 178
+ ],
+ "score": 1.0,
+ "content": ". It is a semidirect product of the scaling group",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 380,
+ 166,
+ 388,
+ 175
+ ],
+ "score": 0.8,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 388,
+ 164,
+ 505,
+ 178
+ ],
+ "score": 1.0,
+ "content": "and the group of translations",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 176,
+ 505,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 176,
+ 136,
+ 186
+ ],
+ "score": 0.88,
+ "content": "T \\cong \\mathbb { R }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 136,
+ 176,
+ 205,
+ 189
+ ],
+ "score": 1.0,
+ "content": ". In other words:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 176,
+ 313,
+ 188
+ ],
+ "score": 0.91,
+ "content": "H = \\{ ( s , t ) | s \\in S , t \\in T \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 313,
+ 176,
+ 505,
+ 189
+ ],
+ "score": 1.0,
+ "content": ". For multiplication of group elements, we have",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 107,
+ 187,
+ 505,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 188,
+ 250,
+ 200
+ ],
+ "score": 0.9,
+ "content": "( s _ { 2 } , t _ { 2 } ) \\cdot ( s _ { 1 } , t _ { 1 } ) = ( s _ { 2 } s _ { 1 } , s _ { 2 } t _ { 1 } + t _ { 2 } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 251,
+ 187,
+ 327,
+ 202
+ ],
+ "score": 1.0,
+ "content": "and for the inverse",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 327,
+ 187,
+ 500,
+ 201
+ ],
+ "score": 0.91,
+ "content": "( s _ { 2 } , t _ { 2 } ) ^ { - 1 } \\cdot ( s _ { 1 } , t _ { 1 } ) = ( s _ { 2 } ^ { - 1 } s _ { 1 } , s _ { 2 } ^ { - 1 } ( t _ { 1 } - t _ { 2 } ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 187,
+ 505,
+ 202
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 104,
+ 198,
+ 505,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 198,
+ 443,
+ 213
+ ],
+ "score": 1.0,
+ "content": "Additionally, for the corresponding scaling and translation transformations, we have",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 443,
+ 200,
+ 505,
+ 211
+ ],
+ "score": 0.88,
+ "content": "L _ { s t } = L _ { s } L _ { t } \\neq",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 210,
+ 347,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 210,
+ 129,
+ 222
+ ],
+ "score": 0.9,
+ "content": "L _ { t } L _ { s }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 129,
+ 210,
+ 347,
+ 223
+ ],
+ "score": 1.0,
+ "content": ", which means that the order of the operations matters.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 7,
+ "bbox_fs": [
+ 104,
+ 121,
+ 506,
+ 223
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 227,
+ 505,
+ 284
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 227,
+ 505,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 227,
+ 421,
+ 239
+ ],
+ "score": 1.0,
+ "content": "From now on, we will work with functions defined on groups, i.e. mappings",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 421,
+ 227,
+ 456,
+ 237
+ ],
+ "score": 0.9,
+ "content": "H \\to \\mathbb { R }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 457,
+ 227,
+ 505,
+ 239
+ ],
+ "score": 1.0,
+ "content": ". Note, that",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 238,
+ 505,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 238,
+ 172,
+ 250
+ ],
+ "score": 1.0,
+ "content": "simple function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 173,
+ 239,
+ 222,
+ 250
+ ],
+ "score": 0.93,
+ "content": "f : \\mathbb { R } \\mathbb { R }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 223,
+ 238,
+ 371,
+ 250
+ ],
+ "score": 1.0,
+ "content": "may be considered as a function on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 372,
+ 239,
+ 382,
+ 248
+ ],
+ "score": 0.8,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 382,
+ 238,
+ 505,
+ 250
+ ],
+ "score": 1.0,
+ "content": "with constant value along the",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 107,
+ 248,
+ 505,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 250,
+ 114,
+ 259
+ ],
+ "score": 0.81,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 115,
+ 248,
+ 340,
+ 262
+ ],
+ "score": 1.0,
+ "content": "axis. Therefore, Equation 4 holds true for functions on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 340,
+ 250,
+ 350,
+ 259
+ ],
+ "score": 0.83,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 351,
+ 248,
+ 505,
+ 262
+ ],
+ "score": 1.0,
+ "content": "as well. One thing we should keep in",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 260,
+ 505,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 260,
+ 222,
+ 272
+ ],
+ "score": 1.0,
+ "content": "mind is that when we apply",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 222,
+ 261,
+ 234,
+ 271
+ ],
+ "score": 0.88,
+ "content": "L _ { s }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 235,
+ 260,
+ 300,
+ 272
+ ],
+ "score": 1.0,
+ "content": "to functions on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 300,
+ 261,
+ 311,
+ 270
+ ],
+ "score": 0.83,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 311,
+ 260,
+ 330,
+ 272
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 330,
+ 261,
+ 338,
+ 270
+ ],
+ "score": 0.83,
+ "content": "\\mathbb { R }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 339,
+ 260,
+ 505,
+ 272
+ ],
+ "score": 1.0,
+ "content": "we use different notations. For example",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 270,
+ 427,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 197,
+ 284
+ ],
+ "score": 0.92,
+ "content": "L _ { s } [ f ] ( x ^ { \\prime } ) = f ( s ^ { - 1 } x ^ { \\prime } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 198,
+ 270,
+ 215,
+ 284
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 216,
+ 271,
+ 427,
+ 284
+ ],
+ "score": 0.83,
+ "content": "\\begin{array} { r } { \\bar { L } _ { s } [ f ] ( s ^ { \\prime } , t ^ { \\prime } ) = f ( ( s , 0 ) ^ { - 1 } ( s ^ { \\prime } , t ^ { \\prime } ) ) = f ( s ^ { - 1 } s ^ { \\prime } , s ^ { - 1 } t ^ { \\prime } ) } \\end{array}",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 14,
+ "bbox_fs": [
+ 106,
+ 227,
+ 505,
+ 284
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 287,
+ 505,
+ 311
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 287,
+ 505,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 287,
+ 304,
+ 301
+ ],
+ "score": 1.0,
+ "content": "Group-Equivariant Convolution Given group",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 304,
+ 288,
+ 313,
+ 298
+ ],
+ "score": 0.82,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 313,
+ 287,
+ 393,
+ 301
+ ],
+ "score": 1.0,
+ "content": "and two functions",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 393,
+ 289,
+ 401,
+ 300
+ ],
+ "score": 0.86,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 401,
+ 287,
+ 421,
+ 301
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 421,
+ 289,
+ 430,
+ 299
+ ],
+ "score": 0.85,
+ "content": "\\psi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 430,
+ 287,
+ 491,
+ 301
+ ],
+ "score": 1.0,
+ "content": "defined on it,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 492,
+ 288,
+ 501,
+ 298
+ ],
+ "score": 0.77,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 287,
+ 505,
+ 301
+ ],
+ "score": 1.0,
+ "content": "-",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 299,
+ 248,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 299,
+ 248,
+ 312
+ ],
+ "score": 1.0,
+ "content": "equivariant convolution is given by",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 17.5,
+ "bbox_fs": [
+ 105,
+ 287,
+ 505,
+ 312
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 165,
+ 313,
+ 445,
+ 340
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 165,
+ 313,
+ 445,
+ 340
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 165,
+ 313,
+ 445,
+ 340
+ ],
+ "score": 0.94,
+ "content": "[ f \\star _ { G } \\psi ] ( g ) = \\int _ { G } f ( g ^ { \\prime } ) L _ { g } [ \\psi ] ( g ^ { \\prime } ) d \\mu ( g ^ { \\prime } ) = \\int _ { G } f ( g ^ { \\prime } ) \\psi ( g ^ { - 1 } g ^ { \\prime } ) d \\mu ( g ^ { \\prime } )",
+ "type": "interline_equation",
+ "image_path": "5988433815b304c1229351695466f4b8fe3fc481fadcd0844c1eef82a7d8bd97.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 19,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 165,
+ 313,
+ 445,
+ 340
+ ],
+ "spans": [],
+ "index": 19
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 344,
+ 505,
+ 389
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 506,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 129,
+ 357
+ ],
+ "score": 1.0,
+ "content": "Here",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 129,
+ 344,
+ 151,
+ 356
+ ],
+ "score": 0.91,
+ "content": "\\mu ( g ^ { \\prime } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 152,
+ 343,
+ 457,
+ 357
+ ],
+ "score": 1.0,
+ "content": "is the Haar measure also known as invariant measure Folland (2016). For",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 457,
+ 344,
+ 489,
+ 354
+ ],
+ "score": 0.91,
+ "content": "T \\cong \\mathbb { R }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 489,
+ 343,
+ 506,
+ 357
+ ],
+ "score": 1.0,
+ "content": "we",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 354,
+ 506,
+ 368
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 354,
+ 127,
+ 368
+ ],
+ "score": 1.0,
+ "content": "have",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 128,
+ 355,
+ 182,
+ 367
+ ],
+ "score": 0.92,
+ "content": "d \\mu ( g ^ { \\prime } ) = d g ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 183,
+ 354,
+ 506,
+ 368
+ ],
+ "score": 1.0,
+ "content": ". For discrete groups, the Haar measure is the counting measure, and integration",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 366,
+ 504,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 497,
+ 379
+ ],
+ "score": 1.0,
+ "content": "becomes a discrete sum. This formula tells us that the output of the convolution evaluated at point",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 498,
+ 370,
+ 504,
+ 378
+ ],
+ "score": 0.73,
+ "content": "g",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 104,
+ 376,
+ 413,
+ 391
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 376,
+ 272,
+ 391
+ ],
+ "score": 1.0,
+ "content": "is the inner product between the function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 272,
+ 378,
+ 279,
+ 389
+ ],
+ "score": 0.86,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 280,
+ 376,
+ 384,
+ 391
+ ],
+ "score": 1.0,
+ "content": "and the transformed filter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 384,
+ 377,
+ 409,
+ 390
+ ],
+ "score": 0.92,
+ "content": "L _ { g } [ \\psi ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 409,
+ 376,
+ 413,
+ 391
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 21.5,
+ "bbox_fs": [
+ 104,
+ 343,
+ 506,
+ 391
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 404,
+ 296,
+ 417
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 402,
+ 298,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 402,
+ 298,
+ 420
+ ],
+ "score": 1.0,
+ "content": "3 SCALE-EQUIVARIANT MAPPINGS",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 428,
+ 384,
+ 441
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 385,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 385,
+ 443
+ ],
+ "score": 1.0,
+ "content": "Now we define the main building blocks of scale-equivariant models.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25,
+ "bbox_fs": [
+ 105,
+ 426,
+ 385,
+ 443
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 446,
+ 505,
+ 491
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 458
+ ],
+ "score": 1.0,
+ "content": "Scale Convolution In order to derive scale convolution, we start from group equivariant convolution",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 457,
+ 505,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 457,
+ 127,
+ 470
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 127,
+ 457,
+ 161,
+ 467
+ ],
+ "score": 0.89,
+ "content": "G = H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 162,
+ 457,
+ 505,
+ 470
+ ],
+ "score": 1.0,
+ "content": ". We first use the property of semidirect product of groups which splits the integral,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 467,
+ 506,
+ 481
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 467,
+ 506,
+ 481
+ ],
+ "score": 1.0,
+ "content": "then choose the appropriate Haar measures, and finally use the properties of steerable filters. Given",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 478,
+ 489,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 157,
+ 492
+ ],
+ "score": 1.0,
+ "content": "the function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 157,
+ 479,
+ 184,
+ 491
+ ],
+ "score": 0.92,
+ "content": "f ( s , t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 185,
+ 478,
+ 269,
+ 492
+ ],
+ "score": 1.0,
+ "content": "and a steerable filter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 269,
+ 479,
+ 302,
+ 491
+ ],
+ "score": 0.93,
+ "content": "\\psi _ { \\sigma } ( s , t )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 302,
+ 478,
+ 348,
+ 492
+ ],
+ "score": 1.0,
+ "content": "defined on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 348,
+ 479,
+ 358,
+ 489
+ ],
+ "score": 0.82,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 358,
+ 478,
+ 489,
+ 492
+ ],
+ "score": 1.0,
+ "content": ", a scale convolution is given by:",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 27.5,
+ "bbox_fs": [
+ 105,
+ 446,
+ 506,
+ 492
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 119,
+ 495,
+ 479,
+ 551
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 495,
+ 479,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 495,
+ 479,
+ 551
+ ],
+ "score": 0.93,
+ "content": "\\begin{array} { l } { { [ f \\star _ { \\cal H } \\psi _ { \\sigma } ] ( s , t ) = \\displaystyle \\int _ { \\cal S } \\int _ { \\cal T } f ( s ^ { \\prime } , t ^ { \\prime } ) { \\cal L } _ { s t } [ \\psi _ { \\sigma } ] ( s ^ { \\prime } , t ^ { \\prime } ) d \\mu ( s ^ { \\prime } ) d \\mu ( t ^ { \\prime } ) } } \\\\ { { { } } } \\\\ { { { } = \\displaystyle \\sum _ { s ^ { \\prime } } \\displaystyle \\int _ { \\cal T } f ( s ^ { \\prime } , t ^ { \\prime } ) \\psi _ { s \\sigma } ( s ^ { - 1 } s ^ { \\prime } , t ^ { \\prime } - t ) d t ^ { \\prime } = \\displaystyle \\sum _ { s ^ { \\prime } } [ f ( s ^ { \\prime } , \\cdot ) \\star \\psi _ { s \\sigma } ( s ^ { - 1 } s ^ { \\prime } , \\cdot ) ] ( t ) } } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "8cb3ab9dc952636da09b958c251b9f4946d4bb834f2be16247217b30e4ccb17c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 31,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 119,
+ 495,
+ 479,
+ 513.6666666666666
+ ],
+ "spans": [],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 119,
+ 513.6666666666666,
+ 479,
+ 532.3333333333333
+ ],
+ "spans": [],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 119,
+ 532.3333333333333,
+ 479,
+ 550.9999999999999
+ ],
+ "spans": [],
+ "index": 32
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 560,
+ 362,
+ 572
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 559,
+ 363,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 559,
+ 185,
+ 573
+ ],
+ "score": 1.0,
+ "content": "And for the case of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 186,
+ 561,
+ 200,
+ 571
+ ],
+ "score": 0.89,
+ "content": "C _ { \\mathrm { i n } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 200,
+ 559,
+ 241,
+ 573
+ ],
+ "score": 1.0,
+ "content": "input and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 241,
+ 561,
+ 259,
+ 571
+ ],
+ "score": 0.9,
+ "content": "C _ { \\mathrm { o u t } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 259,
+ 559,
+ 363,
+ 573
+ ],
+ "score": 1.0,
+ "content": "output channels we have:",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 33,
+ "bbox_fs": [
+ 106,
+ 559,
+ 363,
+ 573
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 146,
+ 576,
+ 465,
+ 610
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 146,
+ 576,
+ 465,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 576,
+ 465,
+ 610
+ ],
+ "score": 0.94,
+ "content": "[ f \\star _ { H } \\psi _ { \\sigma } ] _ { m } ( s , t ) = \\sum _ { n = 1 } ^ { C _ { \\mathrm { i n } } } \\sum _ { s ^ { \\prime } } [ f _ { n } ( s ^ { \\prime } , \\cdot ) \\star \\psi _ { n , m , s \\sigma } ( s ^ { - 1 } s ^ { \\prime } , \\cdot ) ] ( t ) , \\quad m = 1 \\ldots C _ { \\mathrm { o u t } }",
+ "type": "interline_equation",
+ "image_path": "c5aa5fbbeeac0f26cf646a9653a08ff81f5b335b69d41fab9884696d6ccab114.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 35,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 146,
+ 576,
+ 465,
+ 587.3333333333334
+ ],
+ "spans": [],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 146,
+ 587.3333333333334,
+ 465,
+ 598.6666666666667
+ ],
+ "spans": [],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 146,
+ 598.6666666666667,
+ 465,
+ 610.0000000000001
+ ],
+ "spans": [],
+ "index": 36
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 613,
+ 503,
+ 625
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 612,
+ 504,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 612,
+ 398,
+ 626
+ ],
+ "score": 1.0,
+ "content": "The proof of the equivariance of this convolution to transformations from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 398,
+ 614,
+ 408,
+ 623
+ ],
+ "score": 0.82,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 408,
+ 612,
+ 504,
+ 626
+ ],
+ "score": 1.0,
+ "content": "is given in Appendix A.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 37,
+ "bbox_fs": [
+ 106,
+ 612,
+ 504,
+ 626
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 630,
+ 505,
+ 686
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 630,
+ 505,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 630,
+ 505,
+ 642
+ ],
+ "score": 1.0,
+ "content": "Kondor & Trivedi (2018) prove that a feed-forward neural network is equivariant to transformations",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 641,
+ 505,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 641,
+ 129,
+ 653
+ ],
+ "score": 1.0,
+ "content": "from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 129,
+ 642,
+ 138,
+ 651
+ ],
+ "score": 0.81,
+ "content": "G",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 139,
+ 641,
+ 288,
+ 653
+ ],
+ "score": 1.0,
+ "content": "if and only if it is constructed from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 288,
+ 642,
+ 297,
+ 651
+ ],
+ "score": 0.27,
+ "content": "\\mathbf { G }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 297,
+ 641,
+ 505,
+ 653
+ ],
+ "score": 1.0,
+ "content": "-equivariant convolutional layers. Thus Equation 7",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 653,
+ 505,
+ 665
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 653,
+ 505,
+ 665
+ ],
+ "score": 1.0,
+ "content": "shows the most general form of scale-equivariant layers which allows for building scale-equivariant",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 662,
+ 506,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 662,
+ 280,
+ 677
+ ],
+ "score": 1.0,
+ "content": "convolutional networks with such choice of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 280,
+ 664,
+ 288,
+ 673
+ ],
+ "score": 0.75,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 288,
+ 662,
+ 506,
+ 677
+ ],
+ "score": 1.0,
+ "content": ". We will refer to models using scale-equivariant layers",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 673,
+ 426,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 673,
+ 398,
+ 687
+ ],
+ "score": 1.0,
+ "content": "with steerable filters as Scale-Equivariant Steerable Networks, or shortly",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 398,
+ 674,
+ 426,
+ 685
+ ],
+ "score": 0.33,
+ "content": "S \\bar { E } S N ^ { 1 }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 40,
+ "bbox_fs": [
+ 105,
+ 630,
+ 506,
+ 687
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 691,
+ 504,
+ 714
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 690,
+ 506,
+ 704
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 690,
+ 506,
+ 704
+ ],
+ "score": 1.0,
+ "content": "Nonlinearities In order to guarantee the equivariance of the network to scale transformations, we",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 702,
+ 506,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 702,
+ 506,
+ 714
+ ],
+ "score": 1.0,
+ "content": "use scale equivariant nonlinearities. We are free to use simple point-wise nonlinearities. Indeed,",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 104,
+ 81,
+ 424,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 81,
+ 208,
+ 97
+ ],
+ "score": 1.0,
+ "content": "point-wise nonlinearities",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 208,
+ 85,
+ 214,
+ 92
+ ],
+ "score": 0.73,
+ "content": "\\nu",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 215,
+ 81,
+ 424,
+ 97
+ ],
+ "score": 1.0,
+ "content": ", like ReLU, commute with scaling transformations:",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 43.5,
+ "bbox_fs": [
+ 105,
+ 690,
+ 506,
+ 714
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 82,
+ 423,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 81,
+ 424,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 81,
+ 208,
+ 97
+ ],
+ "score": 1.0,
+ "content": "point-wise nonlinearities",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 208,
+ 85,
+ 214,
+ 92
+ ],
+ "score": 0.73,
+ "content": "\\nu",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 215,
+ 81,
+ 424,
+ 97
+ ],
+ "score": 1.0,
+ "content": ", like ReLU, commute with scaling transformations:",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 181,
+ 99,
+ 429,
+ 131
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 181,
+ 99,
+ 429,
+ 131
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 181,
+ 99,
+ 429,
+ 131
+ ],
+ "score": 0.94,
+ "content": "\\begin{array} { r } { [ \\nu \\circ L _ { s } [ f ] ] ( s ^ { \\prime } , x ^ { \\prime } ) = \\nu ( L _ { s } [ f ] ( s ^ { \\prime } , x ^ { \\prime } ) ) = \\nu ( f ( s ^ { - 1 } s ^ { \\prime } , s ^ { - 1 } x ^ { \\prime } ) ) } \\\\ { = \\nu [ f ] ( s ^ { - 1 } s ^ { \\prime } , s ^ { - 1 } x ^ { \\prime } ) = [ L _ { s } \\circ \\nu [ f ] ] ( s ^ { \\prime } , x ^ { \\prime } ) } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "b2eeebeef0c9b259ac75e0c4773c21a6b103e96c44ddf703c1802c70cc02a5a1.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 181,
+ 99,
+ 429,
+ 109.66666666666667
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 181,
+ 109.66666666666667,
+ 429,
+ 120.33333333333334
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 181,
+ 120.33333333333334,
+ 429,
+ 131.0
+ ],
+ "spans": [],
+ "index": 3
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 140,
+ 505,
+ 174
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 141,
+ 505,
+ 153
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 141,
+ 505,
+ 153
+ ],
+ "score": 1.0,
+ "content": "Pooling Until now we did not discuss how to convert an equivariant mapping to invariant one. One",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 152,
+ 506,
+ 164
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 152,
+ 506,
+ 164
+ ],
+ "score": 1.0,
+ "content": "way to do this is to calculate the invariant measure of the signal. In case of translation, such a",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 163,
+ 312,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 163,
+ 312,
+ 175
+ ],
+ "score": 1.0,
+ "content": "measure could be the maximum value for example.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 179,
+ 503,
+ 213
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 178,
+ 505,
+ 192
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 178,
+ 351,
+ 192
+ ],
+ "score": 1.0,
+ "content": "First, we propose the maximum scale projection defined as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 351,
+ 180,
+ 452,
+ 192
+ ],
+ "score": 0.92,
+ "content": "f ( s , x ) \\operatorname* { m a x } _ { s } f ( s , x )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 452,
+ 178,
+ 505,
+ 192
+ ],
+ "score": 1.0,
+ "content": ". This trans-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 190,
+ 505,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 190,
+ 235,
+ 203
+ ],
+ "score": 1.0,
+ "content": "formation projects the function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 191,
+ 243,
+ 202
+ ],
+ "score": 0.86,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 243,
+ 190,
+ 267,
+ 203
+ ],
+ "score": 1.0,
+ "content": "from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 267,
+ 191,
+ 277,
+ 200
+ ],
+ "score": 0.82,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 277,
+ 190,
+ 289,
+ 203
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 290,
+ 191,
+ 298,
+ 200
+ ],
+ "score": 0.81,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 299,
+ 190,
+ 505,
+ 203
+ ],
+ "score": 1.0,
+ "content": ". Therefore, the representation stays equivariant to",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 202,
+ 326,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 202,
+ 326,
+ 213
+ ],
+ "score": 1.0,
+ "content": "scaling, but loses all information about the scale itself.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 218,
+ 505,
+ 252
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 218,
+ 505,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 218,
+ 505,
+ 232
+ ],
+ "score": 1.0,
+ "content": "Second, we are free to use spatial max-pooling with a moving window or global max pooling. Trans-",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 229,
+ 506,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 229,
+ 147,
+ 243
+ ],
+ "score": 1.0,
+ "content": "formation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 148,
+ 229,
+ 246,
+ 241
+ ],
+ "score": 0.91,
+ "content": "f ( s , x ) \\to \\operatorname* { m a x } _ { x } { \\overline { { f } } } ( s , x )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 247,
+ 229,
+ 330,
+ 243
+ ],
+ "score": 1.0,
+ "content": "projects the function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 330,
+ 230,
+ 338,
+ 241
+ ],
+ "score": 0.86,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 338,
+ 229,
+ 359,
+ 243
+ ],
+ "score": 1.0,
+ "content": "from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 360,
+ 230,
+ 370,
+ 239
+ ],
+ "score": 0.79,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 370,
+ 229,
+ 380,
+ 243
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 380,
+ 230,
+ 388,
+ 239
+ ],
+ "score": 0.76,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 389,
+ 229,
+ 506,
+ 243
+ ],
+ "score": 1.0,
+ "content": ". The obtained representation",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 241,
+ 453,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 241,
+ 453,
+ 252
+ ],
+ "score": 1.0,
+ "content": "is invariant to scaling in spatial domain, however, it stores the information about scale.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 257,
+ 505,
+ 302
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 505,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 505,
+ 269
+ ],
+ "score": 1.0,
+ "content": "Finally, we can combine both of these pooling mechanisms in any order. The obtained transforma-",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 268,
+ 505,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 268,
+ 505,
+ 280
+ ],
+ "score": 1.0,
+ "content": "tion produces a scale invariant function. It is useful to utilize this transformation closer to the end of",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 279,
+ 505,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 279,
+ 505,
+ 293
+ ],
+ "score": 1.0,
+ "content": "the network, when the deep representation must be invariant to nuisance input variations, but already",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 290,
+ 237,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 290,
+ 237,
+ 303
+ ],
+ "score": 1.0,
+ "content": "has very rich semantic meaning.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 14.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 318,
+ 219,
+ 330
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 317,
+ 221,
+ 332
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 317,
+ 221,
+ 332
+ ],
+ "score": 1.0,
+ "content": "4 IMPLEMENTATION",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 342,
+ 505,
+ 376
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 504,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 504,
+ 355
+ ],
+ "score": 1.0,
+ "content": "In this paragraph we discuss an efficient implementation of Scale-Equivariant Steerable Networks.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 353,
+ 505,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 353,
+ 505,
+ 366
+ ],
+ "score": 1.0,
+ "content": "We illustrate all algorithms in Figure 1. For simplicity we assume that zero padding is applied when",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 364,
+ 327,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 364,
+ 327,
+ 377
+ ],
+ "score": 1.0,
+ "content": "it is needed for both the spatial axes and the scale axis.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 381,
+ 505,
+ 475
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 381,
+ 505,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 381,
+ 505,
+ 394
+ ],
+ "score": 1.0,
+ "content": "Filter Basis A direct implementation of Equation 7 is impossible due to several limitations. First,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 391,
+ 505,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 391,
+ 236,
+ 406
+ ],
+ "score": 1.0,
+ "content": "the infinite number of scales in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 393,
+ 244,
+ 402
+ ],
+ "score": 0.81,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 245,
+ 391,
+ 505,
+ 406
+ ],
+ "score": 1.0,
+ "content": "calls for a discrete approximation. We truncate the scale group",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 403,
+ 505,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 403,
+ 198,
+ 416
+ ],
+ "score": 1.0,
+ "content": "and limit ourselves to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 199,
+ 404,
+ 213,
+ 415
+ ],
+ "score": 0.89,
+ "content": "N _ { S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 214,
+ 403,
+ 505,
+ 416
+ ],
+ "score": 1.0,
+ "content": "scales and use discrete translations instead of continuous ones. Train-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 104,
+ 414,
+ 505,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 414,
+ 505,
+ 426
+ ],
+ "score": 1.0,
+ "content": "ing of SESN involves searching for the optimal filter in functional space which is a problem",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 425,
+ 505,
+ 438
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 425,
+ 410,
+ 438
+ ],
+ "score": 1.0,
+ "content": "by itself. Rather than solving it directly, we choose a complete basis of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 411,
+ 426,
+ 424,
+ 437
+ ],
+ "score": 0.89,
+ "content": "N _ { b }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 424,
+ 425,
+ 505,
+ 438
+ ],
+ "score": 1.0,
+ "content": "steerable functions",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 107,
+ 433,
+ 507,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 436,
+ 181,
+ 451
+ ],
+ "score": 0.93,
+ "content": "\\dot { \\Psi } = \\{ \\psi _ { s ^ { - 1 } \\sigma , i } \\} _ { i = 1 } ^ { N _ { b } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 182,
+ 433,
+ 507,
+ 454
+ ],
+ "score": 1.0,
+ "content": "and represent convolutional filter as a linear combination of basis functions with",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 104,
+ 447,
+ 507,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 447,
+ 190,
+ 467
+ ],
+ "score": 1.0,
+ "content": "trainable parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 190,
+ 449,
+ 246,
+ 464
+ ],
+ "score": 0.93,
+ "content": "w = \\{ w _ { i } \\} _ { i = 1 } ^ { N _ { b } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 246,
+ 447,
+ 507,
+ 467
+ ],
+ "score": 1.0,
+ "content": ". In other words, we do the following substitution in Equation 7:",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 462,
+ 191,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 462,
+ 191,
+ 475
+ ],
+ "score": 0.88,
+ "content": "\\begin{array} { r } { \\psi _ { \\sigma } \\kappa \\stackrel { - } { = } \\sum _ { i } w _ { i } \\Psi _ { i } } \\end{array}",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 24.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 479,
+ 505,
+ 513
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 506,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 506,
+ 492
+ ],
+ "score": 1.0,
+ "content": "In our experiments we use a basis of 2D Hermite polynomials with 2D Gaussian envelope, as it",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 490,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 490,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "demonstrates good results. The basis is pre-calculated for all scales and fixed. For filters of size",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 107,
+ 499,
+ 488,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 501,
+ 136,
+ 512
+ ],
+ "score": 0.89,
+ "content": "V \\times V",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 136,
+ 499,
+ 294,
+ 515
+ ],
+ "score": 1.0,
+ "content": ", the basis is stored as an array of shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 295,
+ 501,
+ 347,
+ 513
+ ],
+ "score": 0.92,
+ "content": "[ N _ { b } , S , V , V ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 347,
+ 499,
+ 488,
+ 515
+ ],
+ "score": 1.0,
+ "content": ". See Appendix C for more details.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 518,
+ 504,
+ 552
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 516,
+ 505,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 516,
+ 131,
+ 532
+ ],
+ "score": 1.0,
+ "content": "Conv",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 131,
+ 518,
+ 168,
+ 528
+ ],
+ "score": 0.89,
+ "content": "T \\to H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 516,
+ 323,
+ 532
+ ],
+ "score": 1.0,
+ "content": "If the input signal is just a function on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 323,
+ 519,
+ 332,
+ 528
+ ],
+ "score": 0.8,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 332,
+ 516,
+ 399,
+ 532
+ ],
+ "score": 1.0,
+ "content": "with spatial size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 400,
+ 519,
+ 429,
+ 528
+ ],
+ "score": 0.9,
+ "content": "U \\times U",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 429,
+ 516,
+ 505,
+ 532
+ ],
+ "score": 1.0,
+ "content": ", stored as an array",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 529,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 529,
+ 142,
+ 541
+ ],
+ "score": 1.0,
+ "content": "of shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 143,
+ 529,
+ 185,
+ 541
+ ],
+ "score": 0.92,
+ "content": "[ C _ { \\mathrm { i n } } , U , U ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 186,
+ 529,
+ 412,
+ 541
+ ],
+ "score": 1.0,
+ "content": ", then Equation 7 can be simplified. The summation over",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 413,
+ 530,
+ 421,
+ 539
+ ],
+ "score": 0.81,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 421,
+ 529,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": "degenerates, and the",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 540,
+ 301,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 301,
+ 552
+ ],
+ "score": 1.0,
+ "content": "final result can be written in the following form:",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 172,
+ 556,
+ 438,
+ 570
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 172,
+ 556,
+ 438,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 172,
+ 556,
+ 438,
+ 570
+ ],
+ "score": 0.84,
+ "content": "\\mathtt { c o n v T H } ( f , w , \\Psi ) = \\mathtt { s q u e e } z \\in \\left( \\mathtt { c o n v } 2 \\mathrm { d } ( f , \\mathtt { e x p a n d } ( w \\times \\Psi ) ) \\right)",
+ "type": "interline_equation",
+ "image_path": "128a59139f7def3ef1c6918aed47f7ea431e6bba05ddfd59792f60d079b3c1ca.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 35,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 172,
+ 556,
+ 438,
+ 570
+ ],
+ "spans": [],
+ "index": 35
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 580,
+ 505,
+ 636
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 581,
+ 506,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 581,
+ 128,
+ 595
+ ],
+ "score": 1.0,
+ "content": "Here",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 128,
+ 583,
+ 137,
+ 591
+ ],
+ "score": 0.78,
+ "content": "w",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 137,
+ 581,
+ 216,
+ 595
+ ],
+ "score": 1.0,
+ "content": "is an array of shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 217,
+ 581,
+ 273,
+ 593
+ ],
+ "score": 0.92,
+ "content": "[ C _ { \\mathrm { o u t } } , C _ { \\mathrm { i n } } , N _ { b } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 273,
+ 581,
+ 350,
+ 595
+ ],
+ "score": 1.0,
+ "content": ". We compute filter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 350,
+ 582,
+ 376,
+ 592
+ ],
+ "score": 0.91,
+ "content": "w \\times \\Psi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 376,
+ 581,
+ 413,
+ 595
+ ],
+ "score": 1.0,
+ "content": "of shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 413,
+ 581,
+ 487,
+ 594
+ ],
+ "score": 0.9,
+ "content": "[ C _ { \\mathrm { o u t } } , C _ { \\mathrm { i n } } , S , V , V ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 581,
+ 506,
+ 595
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 591,
+ 506,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 591,
+ 183,
+ 605
+ ],
+ "score": 1.0,
+ "content": "expand it to shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 184,
+ 593,
+ 253,
+ 604
+ ],
+ "score": 0.91,
+ "content": "[ C _ { \\mathrm { o u t } } , C _ { \\mathrm { i n } } S , V , V ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 253,
+ 591,
+ 506,
+ 605
+ ],
+ "score": 1.0,
+ "content": ". Then we use standard 2D convolution to produce the output",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 603,
+ 506,
+ 616
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 603,
+ 127,
+ 616
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 127,
+ 603,
+ 152,
+ 614
+ ],
+ "score": 0.91,
+ "content": "C _ { \\mathrm { o u t } } S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 152,
+ 603,
+ 287,
+ 616
+ ],
+ "score": 1.0,
+ "content": "channels and squeeze it to shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 288,
+ 603,
+ 344,
+ 615
+ ],
+ "score": 0.93,
+ "content": "[ C _ { \\mathrm { o u t } } , S , U , U ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 345,
+ 603,
+ 506,
+ 616
+ ],
+ "score": 1.0,
+ "content": ". Note that the output can be viewed as",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 614,
+ 505,
+ 627
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 614,
+ 445,
+ 627
+ ],
+ "score": 1.0,
+ "content": "a stack of feature maps, where all the features in each spatial position are vectors of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 445,
+ 615,
+ 453,
+ 624
+ ],
+ "score": 0.81,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 453,
+ 614,
+ 505,
+ 627
+ ],
+ "score": 1.0,
+ "content": "components",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 624,
+ 289,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 624,
+ 289,
+ 638
+ ],
+ "score": 1.0,
+ "content": "instead of being scalars as in standard CNNs.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 38
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 641,
+ 505,
+ 698
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 640,
+ 505,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 640,
+ 131,
+ 655
+ ],
+ "score": 1.0,
+ "content": "Conv",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 132,
+ 642,
+ 171,
+ 652
+ ],
+ "score": 0.9,
+ "content": "H H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 171,
+ 640,
+ 237,
+ 655
+ ],
+ "score": 1.0,
+ "content": "The function on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 237,
+ 642,
+ 248,
+ 652
+ ],
+ "score": 0.79,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 248,
+ 640,
+ 505,
+ 655
+ ],
+ "score": 1.0,
+ "content": "has a scale axis and therefore there are two options for choosing",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 652,
+ 505,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 652,
+ 505,
+ 666
+ ],
+ "score": 1.0,
+ "content": "weights of the convolutional filter. The filter may have just one scale and, therefore, does not capture",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 664,
+ 505,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 664,
+ 505,
+ 676
+ ],
+ "score": 1.0,
+ "content": "the correlations between different scales of the input function; or, it may have a non-unitary extent",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 107,
+ 675,
+ 505,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 675,
+ 122,
+ 686
+ ],
+ "score": 0.9,
+ "content": "K _ { S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 122,
+ 675,
+ 341,
+ 687
+ ],
+ "score": 1.0,
+ "content": "in the scale axis and capture the correlation between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 341,
+ 675,
+ 356,
+ 686
+ ],
+ "score": 0.89,
+ "content": "K _ { S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 357,
+ 675,
+ 505,
+ 687
+ ],
+ "score": 1.0,
+ "content": "neighboring scales. We refer to the",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 686,
+ 257,
+ 698
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 686,
+ 257,
+ 698
+ ],
+ "score": 1.0,
+ "content": "second case as interscale interaction.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 43
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 102,
+ 702,
+ 493,
+ 715
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 495,
+ 717
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 167,
+ 717
+ ],
+ "score": 1.0,
+ "content": "It the first case",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 167,
+ 705,
+ 176,
+ 713
+ ],
+ "score": 0.76,
+ "content": "w",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 177,
+ 700,
+ 218,
+ 717
+ ],
+ "score": 1.0,
+ "content": "has shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 703,
+ 275,
+ 715
+ ],
+ "score": 0.93,
+ "content": "[ C _ { \\mathrm { o u t } } , C _ { \\mathrm { i n } } , N _ { b } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 276,
+ 700,
+ 495,
+ 717
+ ],
+ "score": 1.0,
+ "content": "and Equation 7 degenerates in the same way as before",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 46
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 149,
+ 719,
+ 460,
+ 733
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 149,
+ 719,
+ 460,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 149,
+ 719,
+ 460,
+ 733
+ ],
+ "score": 0.87,
+ "content": "\\mathtt { c o n v H H } ( f , w , \\Psi ) = \\mathtt { s q u e e } z \\in \\left( \\mathtt { c o n v } 2 \\mathrm { d } ( \\exp \\mathsf { a n d } ( f ) , \\mathtt { e x p a n d } ( w \\times \\Psi ) ) \\right)",
+ "type": "interline_equation",
+ "image_path": "e65cff898868435e3aee58a331636287f343e513a85afceeb1662e2a44444fcb.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 47,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 149,
+ 719,
+ 460,
+ 733
+ ],
+ "spans": [],
+ "index": 47
+ }
+ ]
+ }
+ ],
+ "page_idx": 3,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 26,
+ 294,
+ 38
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 39
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 39
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 759
+ ],
+ "lines": []
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 82,
+ 423,
+ 94
+ ],
+ "lines": [],
+ "index": 0,
+ "bbox_fs": [
+ 104,
+ 81,
+ 424,
+ 97
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 181,
+ 99,
+ 429,
+ 131
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 181,
+ 99,
+ 429,
+ 131
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 181,
+ 99,
+ 429,
+ 131
+ ],
+ "score": 0.94,
+ "content": "\\begin{array} { r } { [ \\nu \\circ L _ { s } [ f ] ] ( s ^ { \\prime } , x ^ { \\prime } ) = \\nu ( L _ { s } [ f ] ( s ^ { \\prime } , x ^ { \\prime } ) ) = \\nu ( f ( s ^ { - 1 } s ^ { \\prime } , s ^ { - 1 } x ^ { \\prime } ) ) } \\\\ { = \\nu [ f ] ( s ^ { - 1 } s ^ { \\prime } , s ^ { - 1 } x ^ { \\prime } ) = [ L _ { s } \\circ \\nu [ f ] ] ( s ^ { \\prime } , x ^ { \\prime } ) } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "b2eeebeef0c9b259ac75e0c4773c21a6b103e96c44ddf703c1802c70cc02a5a1.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 181,
+ 99,
+ 429,
+ 109.66666666666667
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 181,
+ 109.66666666666667,
+ 429,
+ 120.33333333333334
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 181,
+ 120.33333333333334,
+ 429,
+ 131.0
+ ],
+ "spans": [],
+ "index": 3
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 140,
+ 505,
+ 174
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 141,
+ 505,
+ 153
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 141,
+ 505,
+ 153
+ ],
+ "score": 1.0,
+ "content": "Pooling Until now we did not discuss how to convert an equivariant mapping to invariant one. One",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 152,
+ 506,
+ 164
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 152,
+ 506,
+ 164
+ ],
+ "score": 1.0,
+ "content": "way to do this is to calculate the invariant measure of the signal. In case of translation, such a",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 163,
+ 312,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 163,
+ 312,
+ 175
+ ],
+ "score": 1.0,
+ "content": "measure could be the maximum value for example.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5,
+ "bbox_fs": [
+ 105,
+ 141,
+ 506,
+ 175
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 179,
+ 503,
+ 213
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 178,
+ 505,
+ 192
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 178,
+ 351,
+ 192
+ ],
+ "score": 1.0,
+ "content": "First, we propose the maximum scale projection defined as",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 351,
+ 180,
+ 452,
+ 192
+ ],
+ "score": 0.92,
+ "content": "f ( s , x ) \\operatorname* { m a x } _ { s } f ( s , x )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 452,
+ 178,
+ 505,
+ 192
+ ],
+ "score": 1.0,
+ "content": ". This trans-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 190,
+ 505,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 190,
+ 235,
+ 203
+ ],
+ "score": 1.0,
+ "content": "formation projects the function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 235,
+ 191,
+ 243,
+ 202
+ ],
+ "score": 0.86,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 243,
+ 190,
+ 267,
+ 203
+ ],
+ "score": 1.0,
+ "content": "from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 267,
+ 191,
+ 277,
+ 200
+ ],
+ "score": 0.82,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 277,
+ 190,
+ 289,
+ 203
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 290,
+ 191,
+ 298,
+ 200
+ ],
+ "score": 0.81,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 299,
+ 190,
+ 505,
+ 203
+ ],
+ "score": 1.0,
+ "content": ". Therefore, the representation stays equivariant to",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 202,
+ 326,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 202,
+ 326,
+ 213
+ ],
+ "score": 1.0,
+ "content": "scaling, but loses all information about the scale itself.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8,
+ "bbox_fs": [
+ 105,
+ 178,
+ 505,
+ 213
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 218,
+ 505,
+ 252
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 218,
+ 505,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 218,
+ 505,
+ 232
+ ],
+ "score": 1.0,
+ "content": "Second, we are free to use spatial max-pooling with a moving window or global max pooling. Trans-",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 229,
+ 506,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 229,
+ 147,
+ 243
+ ],
+ "score": 1.0,
+ "content": "formation",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 148,
+ 229,
+ 246,
+ 241
+ ],
+ "score": 0.91,
+ "content": "f ( s , x ) \\to \\operatorname* { m a x } _ { x } { \\overline { { f } } } ( s , x )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 247,
+ 229,
+ 330,
+ 243
+ ],
+ "score": 1.0,
+ "content": "projects the function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 330,
+ 230,
+ 338,
+ 241
+ ],
+ "score": 0.86,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 338,
+ 229,
+ 359,
+ 243
+ ],
+ "score": 1.0,
+ "content": "from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 360,
+ 230,
+ 370,
+ 239
+ ],
+ "score": 0.79,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 370,
+ 229,
+ 380,
+ 243
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 380,
+ 230,
+ 388,
+ 239
+ ],
+ "score": 0.76,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 389,
+ 229,
+ 506,
+ 243
+ ],
+ "score": 1.0,
+ "content": ". The obtained representation",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 241,
+ 453,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 241,
+ 453,
+ 252
+ ],
+ "score": 1.0,
+ "content": "is invariant to scaling in spatial domain, however, it stores the information about scale.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11,
+ "bbox_fs": [
+ 105,
+ 218,
+ 506,
+ 252
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 257,
+ 505,
+ 302
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 505,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 505,
+ 269
+ ],
+ "score": 1.0,
+ "content": "Finally, we can combine both of these pooling mechanisms in any order. The obtained transforma-",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 268,
+ 505,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 268,
+ 505,
+ 280
+ ],
+ "score": 1.0,
+ "content": "tion produces a scale invariant function. It is useful to utilize this transformation closer to the end of",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 279,
+ 505,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 279,
+ 505,
+ 293
+ ],
+ "score": 1.0,
+ "content": "the network, when the deep representation must be invariant to nuisance input variations, but already",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 290,
+ 237,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 290,
+ 237,
+ 303
+ ],
+ "score": 1.0,
+ "content": "has very rich semantic meaning.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 14.5,
+ "bbox_fs": [
+ 105,
+ 258,
+ 505,
+ 303
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 318,
+ 219,
+ 330
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 317,
+ 221,
+ 332
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 317,
+ 221,
+ 332
+ ],
+ "score": 1.0,
+ "content": "4 IMPLEMENTATION",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 342,
+ 505,
+ 376
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 504,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 504,
+ 355
+ ],
+ "score": 1.0,
+ "content": "In this paragraph we discuss an efficient implementation of Scale-Equivariant Steerable Networks.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 353,
+ 505,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 353,
+ 505,
+ 366
+ ],
+ "score": 1.0,
+ "content": "We illustrate all algorithms in Figure 1. For simplicity we assume that zero padding is applied when",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 364,
+ 327,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 364,
+ 327,
+ 377
+ ],
+ "score": 1.0,
+ "content": "it is needed for both the spatial axes and the scale axis.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 19,
+ "bbox_fs": [
+ 105,
+ 343,
+ 505,
+ 377
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 381,
+ 505,
+ 475
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 381,
+ 505,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 381,
+ 505,
+ 394
+ ],
+ "score": 1.0,
+ "content": "Filter Basis A direct implementation of Equation 7 is impossible due to several limitations. First,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 391,
+ 505,
+ 406
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 391,
+ 236,
+ 406
+ ],
+ "score": 1.0,
+ "content": "the infinite number of scales in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 393,
+ 244,
+ 402
+ ],
+ "score": 0.81,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 245,
+ 391,
+ 505,
+ 406
+ ],
+ "score": 1.0,
+ "content": "calls for a discrete approximation. We truncate the scale group",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 403,
+ 505,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 403,
+ 198,
+ 416
+ ],
+ "score": 1.0,
+ "content": "and limit ourselves to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 199,
+ 404,
+ 213,
+ 415
+ ],
+ "score": 0.89,
+ "content": "N _ { S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 214,
+ 403,
+ 505,
+ 416
+ ],
+ "score": 1.0,
+ "content": "scales and use discrete translations instead of continuous ones. Train-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 104,
+ 414,
+ 505,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 414,
+ 505,
+ 426
+ ],
+ "score": 1.0,
+ "content": "ing of SESN involves searching for the optimal filter in functional space which is a problem",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 425,
+ 505,
+ 438
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 425,
+ 410,
+ 438
+ ],
+ "score": 1.0,
+ "content": "by itself. Rather than solving it directly, we choose a complete basis of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 411,
+ 426,
+ 424,
+ 437
+ ],
+ "score": 0.89,
+ "content": "N _ { b }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 424,
+ 425,
+ 505,
+ 438
+ ],
+ "score": 1.0,
+ "content": "steerable functions",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 107,
+ 433,
+ 507,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 436,
+ 181,
+ 451
+ ],
+ "score": 0.93,
+ "content": "\\dot { \\Psi } = \\{ \\psi _ { s ^ { - 1 } \\sigma , i } \\} _ { i = 1 } ^ { N _ { b } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 182,
+ 433,
+ 507,
+ 454
+ ],
+ "score": 1.0,
+ "content": "and represent convolutional filter as a linear combination of basis functions with",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 104,
+ 447,
+ 507,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 447,
+ 190,
+ 467
+ ],
+ "score": 1.0,
+ "content": "trainable parameters",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 190,
+ 449,
+ 246,
+ 464
+ ],
+ "score": 0.93,
+ "content": "w = \\{ w _ { i } \\} _ { i = 1 } ^ { N _ { b } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 246,
+ 447,
+ 507,
+ 467
+ ],
+ "score": 1.0,
+ "content": ". In other words, we do the following substitution in Equation 7:",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 462,
+ 191,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 462,
+ 191,
+ 475
+ ],
+ "score": 0.88,
+ "content": "\\begin{array} { r } { \\psi _ { \\sigma } \\kappa \\stackrel { - } { = } \\sum _ { i } w _ { i } \\Psi _ { i } } \\end{array}",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 24.5,
+ "bbox_fs": [
+ 104,
+ 381,
+ 507,
+ 475
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 479,
+ 505,
+ 513
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 506,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 506,
+ 492
+ ],
+ "score": 1.0,
+ "content": "In our experiments we use a basis of 2D Hermite polynomials with 2D Gaussian envelope, as it",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 490,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 490,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "demonstrates good results. The basis is pre-calculated for all scales and fixed. For filters of size",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 107,
+ 499,
+ 488,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 501,
+ 136,
+ 512
+ ],
+ "score": 0.89,
+ "content": "V \\times V",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 136,
+ 499,
+ 294,
+ 515
+ ],
+ "score": 1.0,
+ "content": ", the basis is stored as an array of shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 295,
+ 501,
+ 347,
+ 513
+ ],
+ "score": 0.92,
+ "content": "[ N _ { b } , S , V , V ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 347,
+ 499,
+ 488,
+ 515
+ ],
+ "score": 1.0,
+ "content": ". See Appendix C for more details.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 105,
+ 478,
+ 506,
+ 515
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 518,
+ 504,
+ 552
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 516,
+ 505,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 516,
+ 131,
+ 532
+ ],
+ "score": 1.0,
+ "content": "Conv",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 131,
+ 518,
+ 168,
+ 528
+ ],
+ "score": 0.89,
+ "content": "T \\to H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 168,
+ 516,
+ 323,
+ 532
+ ],
+ "score": 1.0,
+ "content": "If the input signal is just a function on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 323,
+ 519,
+ 332,
+ 528
+ ],
+ "score": 0.8,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 332,
+ 516,
+ 399,
+ 532
+ ],
+ "score": 1.0,
+ "content": "with spatial size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 400,
+ 519,
+ 429,
+ 528
+ ],
+ "score": 0.9,
+ "content": "U \\times U",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 429,
+ 516,
+ 505,
+ 532
+ ],
+ "score": 1.0,
+ "content": ", stored as an array",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 529,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 529,
+ 142,
+ 541
+ ],
+ "score": 1.0,
+ "content": "of shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 143,
+ 529,
+ 185,
+ 541
+ ],
+ "score": 0.92,
+ "content": "[ C _ { \\mathrm { i n } } , U , U ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 186,
+ 529,
+ 412,
+ 541
+ ],
+ "score": 1.0,
+ "content": ", then Equation 7 can be simplified. The summation over",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 413,
+ 530,
+ 421,
+ 539
+ ],
+ "score": 0.81,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 421,
+ 529,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": "degenerates, and the",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 540,
+ 301,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 301,
+ 552
+ ],
+ "score": 1.0,
+ "content": "final result can be written in the following form:",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 33,
+ "bbox_fs": [
+ 105,
+ 516,
+ 505,
+ 552
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 172,
+ 556,
+ 438,
+ 570
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 172,
+ 556,
+ 438,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 172,
+ 556,
+ 438,
+ 570
+ ],
+ "score": 0.84,
+ "content": "\\mathtt { c o n v T H } ( f , w , \\Psi ) = \\mathtt { s q u e e } z \\in \\left( \\mathtt { c o n v } 2 \\mathrm { d } ( f , \\mathtt { e x p a n d } ( w \\times \\Psi ) ) \\right)",
+ "type": "interline_equation",
+ "image_path": "128a59139f7def3ef1c6918aed47f7ea431e6bba05ddfd59792f60d079b3c1ca.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 35,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 172,
+ 556,
+ 438,
+ 570
+ ],
+ "spans": [],
+ "index": 35
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 580,
+ 505,
+ 636
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 581,
+ 506,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 581,
+ 128,
+ 595
+ ],
+ "score": 1.0,
+ "content": "Here",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 128,
+ 583,
+ 137,
+ 591
+ ],
+ "score": 0.78,
+ "content": "w",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 137,
+ 581,
+ 216,
+ 595
+ ],
+ "score": 1.0,
+ "content": "is an array of shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 217,
+ 581,
+ 273,
+ 593
+ ],
+ "score": 0.92,
+ "content": "[ C _ { \\mathrm { o u t } } , C _ { \\mathrm { i n } } , N _ { b } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 273,
+ 581,
+ 350,
+ 595
+ ],
+ "score": 1.0,
+ "content": ". We compute filter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 350,
+ 582,
+ 376,
+ 592
+ ],
+ "score": 0.91,
+ "content": "w \\times \\Psi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 376,
+ 581,
+ 413,
+ 595
+ ],
+ "score": 1.0,
+ "content": "of shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 413,
+ 581,
+ 487,
+ 594
+ ],
+ "score": 0.9,
+ "content": "[ C _ { \\mathrm { o u t } } , C _ { \\mathrm { i n } } , S , V , V ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 581,
+ 506,
+ 595
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 591,
+ 506,
+ 605
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 591,
+ 183,
+ 605
+ ],
+ "score": 1.0,
+ "content": "expand it to shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 184,
+ 593,
+ 253,
+ 604
+ ],
+ "score": 0.91,
+ "content": "[ C _ { \\mathrm { o u t } } , C _ { \\mathrm { i n } } S , V , V ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 253,
+ 591,
+ 506,
+ 605
+ ],
+ "score": 1.0,
+ "content": ". Then we use standard 2D convolution to produce the output",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 603,
+ 506,
+ 616
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 603,
+ 127,
+ 616
+ ],
+ "score": 1.0,
+ "content": "with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 127,
+ 603,
+ 152,
+ 614
+ ],
+ "score": 0.91,
+ "content": "C _ { \\mathrm { o u t } } S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 152,
+ 603,
+ 287,
+ 616
+ ],
+ "score": 1.0,
+ "content": "channels and squeeze it to shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 288,
+ 603,
+ 344,
+ 615
+ ],
+ "score": 0.93,
+ "content": "[ C _ { \\mathrm { o u t } } , S , U , U ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 345,
+ 603,
+ 506,
+ 616
+ ],
+ "score": 1.0,
+ "content": ". Note that the output can be viewed as",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 614,
+ 505,
+ 627
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 614,
+ 445,
+ 627
+ ],
+ "score": 1.0,
+ "content": "a stack of feature maps, where all the features in each spatial position are vectors of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 445,
+ 615,
+ 453,
+ 624
+ ],
+ "score": 0.81,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 453,
+ 614,
+ 505,
+ 627
+ ],
+ "score": 1.0,
+ "content": "components",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 624,
+ 289,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 624,
+ 289,
+ 638
+ ],
+ "score": 1.0,
+ "content": "instead of being scalars as in standard CNNs.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 38,
+ "bbox_fs": [
+ 105,
+ 581,
+ 506,
+ 638
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 641,
+ 505,
+ 698
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 640,
+ 505,
+ 655
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 640,
+ 131,
+ 655
+ ],
+ "score": 1.0,
+ "content": "Conv",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 132,
+ 642,
+ 171,
+ 652
+ ],
+ "score": 0.9,
+ "content": "H H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 171,
+ 640,
+ 237,
+ 655
+ ],
+ "score": 1.0,
+ "content": "The function on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 237,
+ 642,
+ 248,
+ 652
+ ],
+ "score": 0.79,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 248,
+ 640,
+ 505,
+ 655
+ ],
+ "score": 1.0,
+ "content": "has a scale axis and therefore there are two options for choosing",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 652,
+ 505,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 652,
+ 505,
+ 666
+ ],
+ "score": 1.0,
+ "content": "weights of the convolutional filter. The filter may have just one scale and, therefore, does not capture",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 664,
+ 505,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 664,
+ 505,
+ 676
+ ],
+ "score": 1.0,
+ "content": "the correlations between different scales of the input function; or, it may have a non-unitary extent",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 107,
+ 675,
+ 505,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 675,
+ 122,
+ 686
+ ],
+ "score": 0.9,
+ "content": "K _ { S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 122,
+ 675,
+ 341,
+ 687
+ ],
+ "score": 1.0,
+ "content": "in the scale axis and capture the correlation between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 341,
+ 675,
+ 356,
+ 686
+ ],
+ "score": 0.89,
+ "content": "K _ { S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 357,
+ 675,
+ 505,
+ 687
+ ],
+ "score": 1.0,
+ "content": "neighboring scales. We refer to the",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 686,
+ 257,
+ 698
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 686,
+ 257,
+ 698
+ ],
+ "score": 1.0,
+ "content": "second case as interscale interaction.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 43,
+ "bbox_fs": [
+ 105,
+ 640,
+ 505,
+ 698
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 102,
+ 702,
+ 493,
+ 715
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 495,
+ 717
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 700,
+ 167,
+ 717
+ ],
+ "score": 1.0,
+ "content": "It the first case",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 167,
+ 705,
+ 176,
+ 713
+ ],
+ "score": 0.76,
+ "content": "w",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 177,
+ 700,
+ 218,
+ 717
+ ],
+ "score": 1.0,
+ "content": "has shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 703,
+ 275,
+ 715
+ ],
+ "score": 0.93,
+ "content": "[ C _ { \\mathrm { o u t } } , C _ { \\mathrm { i n } } , N _ { b } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 276,
+ 700,
+ 495,
+ 717
+ ],
+ "score": 1.0,
+ "content": "and Equation 7 degenerates in the same way as before",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 46,
+ "bbox_fs": [
+ 105,
+ 700,
+ 495,
+ 717
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 149,
+ 719,
+ 460,
+ 733
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 149,
+ 719,
+ 460,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 149,
+ 719,
+ 460,
+ 733
+ ],
+ "score": 0.87,
+ "content": "\\mathtt { c o n v H H } ( f , w , \\Psi ) = \\mathtt { s q u e e } z \\in \\left( \\mathtt { c o n v } 2 \\mathrm { d } ( \\exp \\mathsf { a n d } ( f ) , \\mathtt { e x p a n d } ( w \\times \\Psi ) ) \\right)",
+ "type": "interline_equation",
+ "image_path": "e65cff898868435e3aee58a331636287f343e513a85afceeb1662e2a44444fcb.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 47,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 149,
+ 719,
+ 460,
+ 733
+ ],
+ "spans": [],
+ "index": 47
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 82,
+ 504,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 81,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 81,
+ 153,
+ 96
+ ],
+ "score": 1.0,
+ "content": "We expand",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 153,
+ 83,
+ 160,
+ 94
+ ],
+ "score": 0.85,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 160,
+ 81,
+ 242,
+ 96
+ ],
+ "score": 1.0,
+ "content": "to an array of shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 242,
+ 82,
+ 291,
+ 95
+ ],
+ "score": 0.93,
+ "content": "[ C _ { \\mathrm { i n } } S , U , U ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 291,
+ 81,
+ 340,
+ 96
+ ],
+ "score": 1.0,
+ "content": "and expand",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 340,
+ 83,
+ 368,
+ 93
+ ],
+ "score": 0.91,
+ "content": "w \\times \\Psi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 368,
+ 81,
+ 425,
+ 96
+ ],
+ "score": 1.0,
+ "content": "to have shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 425,
+ 82,
+ 501,
+ 95
+ ],
+ "score": 0.9,
+ "content": "[ C _ { \\mathrm { o u t } } S , C _ { \\mathrm { i n } } S , V , V ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 502,
+ 81,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 399,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 399,
+ 106
+ ],
+ "score": 1.0,
+ "content": "The result of the convolution is then squeezed in the same way as before.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 110,
+ 504,
+ 144
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 506,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 254,
+ 123
+ ],
+ "score": 1.0,
+ "content": "In the case of interscale interaction,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 255,
+ 113,
+ 264,
+ 121
+ ],
+ "score": 0.7,
+ "content": "w",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 264,
+ 110,
+ 308,
+ 123
+ ],
+ "score": 1.0,
+ "content": "has shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 308,
+ 110,
+ 383,
+ 122
+ ],
+ "score": 0.92,
+ "content": "[ C _ { \\mathrm { o u t } } , C _ { \\mathrm { i n } } , K _ { S } , N _ { b } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 384,
+ 110,
+ 506,
+ 123
+ ],
+ "score": 1.0,
+ "content": ". We iterate over all scales in",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 121,
+ 505,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 121,
+ 175,
+ 134
+ ],
+ "score": 1.0,
+ "content": "interaction, shift",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 176,
+ 122,
+ 183,
+ 133
+ ],
+ "score": 0.85,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 183,
+ 121,
+ 374,
+ 134
+ ],
+ "score": 1.0,
+ "content": "for each scale, choose a corresponding part of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 375,
+ 124,
+ 383,
+ 131
+ ],
+ "score": 0.76,
+ "content": "w",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 383,
+ 121,
+ 505,
+ 134
+ ],
+ "score": 1.0,
+ "content": ", and apply convHH to them.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 132,
+ 285,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 132,
+ 192,
+ 144
+ ],
+ "score": 1.0,
+ "content": "We sum the obtained",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 193,
+ 133,
+ 208,
+ 144
+ ],
+ "score": 0.9,
+ "content": "K _ { S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 208,
+ 132,
+ 285,
+ 144
+ ],
+ "score": 1.0,
+ "content": "results afterwards.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 138,
+ 168,
+ 476,
+ 303
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 138,
+ 168,
+ 476,
+ 303
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 168,
+ 476,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 168,
+ 476,
+ 303
+ ],
+ "score": 0.915,
+ "type": "image",
+ "image_path": "3a842654dbcbf77a7b84f924b43c512bc70871657b145e3b30693fa35d2dfeda.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 138,
+ 168,
+ 476,
+ 213.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 138,
+ 213.0,
+ 476,
+ 258.0
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 138,
+ 258.0,
+ 476,
+ 303.0
+ ],
+ "spans": [],
+ "index": 7
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 326,
+ 505,
+ 382
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 327,
+ 506,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 327,
+ 506,
+ 339
+ ],
+ "score": 1.0,
+ "content": "Figure 1: Left: the way steerable filters are computed using steerable filter basis. Middle and right: a",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 336,
+ 506,
+ 351
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 336,
+ 506,
+ 351
+ ],
+ "score": 1.0,
+ "content": "representation of scale-convolution using Equation 9 and Equation 10. As an example we use input",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 347,
+ 504,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 347,
+ 132,
+ 361
+ ],
+ "score": 1.0,
+ "content": "signal",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 133,
+ 349,
+ 140,
+ 360
+ ],
+ "score": 0.86,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 140,
+ 347,
+ 272,
+ 361
+ ],
+ "score": 1.0,
+ "content": "with 3 channels. It has 1 scale on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 272,
+ 349,
+ 281,
+ 358
+ ],
+ "score": 0.75,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 281,
+ 347,
+ 343,
+ 361
+ ],
+ "score": 1.0,
+ "content": "and 4 scales on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 349,
+ 354,
+ 358
+ ],
+ "score": 0.74,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 354,
+ 347,
+ 459,
+ 361
+ ],
+ "score": 1.0,
+ "content": ". It is convolved with filter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 459,
+ 349,
+ 504,
+ 359
+ ],
+ "score": 0.9,
+ "content": "\\kappa = w \\times \\Psi",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 360,
+ 505,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 360,
+ 505,
+ 371
+ ],
+ "score": 1.0,
+ "content": "without scale interaction, which produces the output with 2 channels and 4 scales as well. Here we",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 370,
+ 497,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 370,
+ 497,
+ 383
+ ],
+ "score": 1.0,
+ "content": "represent only channels of the signals and the filter. Spatial components are hidden for simplicity.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 8.0
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 408,
+ 211,
+ 421
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 213,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 213,
+ 423
+ ],
+ "score": 1.0,
+ "content": "5 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 434,
+ 505,
+ 566
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 434,
+ 506,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 434,
+ 506,
+ 448
+ ],
+ "score": 1.0,
+ "content": "Various works on group-equivariant convolutional networks have been published recently. These",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 458
+ ],
+ "score": 1.0,
+ "content": "works have considered roto-translation groups in 2D Cohen & Welling (2016a); Hoogeboom et al.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 457,
+ 505,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 457,
+ 505,
+ 468
+ ],
+ "score": 1.0,
+ "content": "(2018); Worrall et al. (2017); Weiler & Cesa (2019) and 3D Worrall & Brostow (2018); Kondor",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 466,
+ 506,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 506,
+ 480
+ ],
+ "score": 1.0,
+ "content": "(2018); Thomas et al. (2018) and rotation equivariant networks in 3D Cohen et al. (2017); Esteves",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 478,
+ 505,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 505,
+ 491
+ ],
+ "score": 1.0,
+ "content": "et al. (2018); Cohen et al. (2019). In Freeman & Adelson (1991) authors describe the algorithm",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 104,
+ 488,
+ 506,
+ 503
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 488,
+ 506,
+ 503
+ ],
+ "score": 1.0,
+ "content": "for designing steerable filters for rotations. Rotation steerable filters are used in Cohen & Welling",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 513
+ ],
+ "score": 1.0,
+ "content": "(2016b); Weiler et al. (2018a;b) for building equivariant networks. In Jacobsen et al. (2017) the",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 510,
+ 506,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 510,
+ 369,
+ 525
+ ],
+ "score": 1.0,
+ "content": "authors build convolutional blocks locally equivariant to arbitrary",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 369,
+ 512,
+ 376,
+ 521
+ ],
+ "score": 0.81,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 376,
+ 510,
+ 506,
+ 525
+ ],
+ "score": 1.0,
+ "content": "-parameter Lie group by using a",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 535
+ ],
+ "score": 1.0,
+ "content": "steerable basis. And in Murugan et al. the authors discuss the approach for learning steerable filters",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 104,
+ 532,
+ 506,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 532,
+ 506,
+ 546
+ ],
+ "score": 1.0,
+ "content": "from data. To date, the majority of papers on group equivariant networks have considered rotations",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 104,
+ 543,
+ 506,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 543,
+ 506,
+ 558
+ ],
+ "score": 1.0,
+ "content": "in 2D and 3D, but have not payed attention to scale symmetry. As we have argued above, it is a",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 225,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 225,
+ 568
+ ],
+ "score": 1.0,
+ "content": "fundamentally different case.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 19.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 572,
+ 505,
+ 649
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 572,
+ 505,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 572,
+ 505,
+ 585
+ ],
+ "score": 1.0,
+ "content": "Many papers and even conferences have been dedicated to image scale-space — a concept where",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 583,
+ 505,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 583,
+ 505,
+ 595
+ ],
+ "score": 1.0,
+ "content": "the image is analyzed together with all its downscaled versions. Initially introduced in Iijima (1959)",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 607
+ ],
+ "score": 1.0,
+ "content": "and later developed by Witkin (1987); Perona & Malik (1990); Lindeberg (2013) scale space relies",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "on the scale symmetry of images. The differential structure of the image Koenderink (1984) allows",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "one to make a connection between image formation mechanisms and the space of solutions of the 2-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 641
+ ],
+ "score": 1.0,
+ "content": "dimensional heat equation, which significantly improved the image analysis models in the pre-deep",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 159,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 159,
+ 652
+ ],
+ "score": 1.0,
+ "content": "learning era.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 654,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 654,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 654,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "One of the first works on scale equivariance and local scale invariance in the framework of CNNs was",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 677
+ ],
+ "score": 1.0,
+ "content": "proposed by Xu et al. (2014) named SiCNN. The authors describe the model with siamese CNNs,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "where the filters of each instance are rescaled using interpolation techniques. This is the simplest",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "case of equivariance where no interaction between different scales is done in intermediate layers. In",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "SI-ConvNet by Kanazawa et al. (2014) the original network is modified such that, in each layer, the",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "input is first rescaled, then convolved and rescaled back to the original size. Finally, the response",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 720,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 720,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "with the maximum values is chosen between the scales. Thus, the model is locally scale-invariant. In",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "page_idx": 4,
+ "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 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 763
+ ],
+ "score": 1.0,
+ "content": "5",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "list",
+ "bbox": [
+ 104,
+ 82,
+ 504,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 81,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 81,
+ 153,
+ 96
+ ],
+ "score": 1.0,
+ "content": "We expand",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 153,
+ 83,
+ 160,
+ 94
+ ],
+ "score": 0.85,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 160,
+ 81,
+ 242,
+ 96
+ ],
+ "score": 1.0,
+ "content": "to an array of shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 242,
+ 82,
+ 291,
+ 95
+ ],
+ "score": 0.93,
+ "content": "[ C _ { \\mathrm { i n } } S , U , U ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 291,
+ 81,
+ 340,
+ 96
+ ],
+ "score": 1.0,
+ "content": "and expand",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 340,
+ 83,
+ 368,
+ 93
+ ],
+ "score": 0.91,
+ "content": "w \\times \\Psi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 368,
+ 81,
+ 425,
+ 96
+ ],
+ "score": 1.0,
+ "content": "to have shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 425,
+ 82,
+ 501,
+ 95
+ ],
+ "score": 0.9,
+ "content": "[ C _ { \\mathrm { o u t } } S , C _ { \\mathrm { i n } } S , V , V ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 502,
+ 81,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 0,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 93,
+ 399,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 93,
+ 399,
+ 106
+ ],
+ "score": 1.0,
+ "content": "The result of the convolution is then squeezed in the same way as before.",
+ "type": "text"
+ }
+ ],
+ "index": 1,
+ "is_list_start_line": true,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 0.5,
+ "bbox_fs": [
+ 106,
+ 81,
+ 505,
+ 106
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 110,
+ 504,
+ 144
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 506,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 254,
+ 123
+ ],
+ "score": 1.0,
+ "content": "In the case of interscale interaction,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 255,
+ 113,
+ 264,
+ 121
+ ],
+ "score": 0.7,
+ "content": "w",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 264,
+ 110,
+ 308,
+ 123
+ ],
+ "score": 1.0,
+ "content": "has shape",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 308,
+ 110,
+ 383,
+ 122
+ ],
+ "score": 0.92,
+ "content": "[ C _ { \\mathrm { o u t } } , C _ { \\mathrm { i n } } , K _ { S } , N _ { b } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 384,
+ 110,
+ 506,
+ 123
+ ],
+ "score": 1.0,
+ "content": ". We iterate over all scales in",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 121,
+ 505,
+ 134
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 121,
+ 175,
+ 134
+ ],
+ "score": 1.0,
+ "content": "interaction, shift",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 176,
+ 122,
+ 183,
+ 133
+ ],
+ "score": 0.85,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 183,
+ 121,
+ 374,
+ 134
+ ],
+ "score": 1.0,
+ "content": "for each scale, choose a corresponding part of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 375,
+ 124,
+ 383,
+ 131
+ ],
+ "score": 0.76,
+ "content": "w",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 383,
+ 121,
+ 505,
+ 134
+ ],
+ "score": 1.0,
+ "content": ", and apply convHH to them.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 132,
+ 285,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 132,
+ 192,
+ 144
+ ],
+ "score": 1.0,
+ "content": "We sum the obtained",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 193,
+ 133,
+ 208,
+ 144
+ ],
+ "score": 0.9,
+ "content": "K _ { S }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 208,
+ 132,
+ 285,
+ 144
+ ],
+ "score": 1.0,
+ "content": "results afterwards.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3,
+ "bbox_fs": [
+ 105,
+ 110,
+ 506,
+ 144
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 138,
+ 168,
+ 476,
+ 303
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 138,
+ 168,
+ 476,
+ 303
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 168,
+ 476,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 168,
+ 476,
+ 303
+ ],
+ "score": 0.915,
+ "type": "image",
+ "image_path": "3a842654dbcbf77a7b84f924b43c512bc70871657b145e3b30693fa35d2dfeda.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 138,
+ 168,
+ 476,
+ 213.0
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 138,
+ 213.0,
+ 476,
+ 258.0
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 138,
+ 258.0,
+ 476,
+ 303.0
+ ],
+ "spans": [],
+ "index": 7
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 326,
+ 505,
+ 382
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 327,
+ 506,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 327,
+ 506,
+ 339
+ ],
+ "score": 1.0,
+ "content": "Figure 1: Left: the way steerable filters are computed using steerable filter basis. Middle and right: a",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 336,
+ 506,
+ 351
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 336,
+ 506,
+ 351
+ ],
+ "score": 1.0,
+ "content": "representation of scale-convolution using Equation 9 and Equation 10. As an example we use input",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 347,
+ 504,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 347,
+ 132,
+ 361
+ ],
+ "score": 1.0,
+ "content": "signal",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 133,
+ 349,
+ 140,
+ 360
+ ],
+ "score": 0.86,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 140,
+ 347,
+ 272,
+ 361
+ ],
+ "score": 1.0,
+ "content": "with 3 channels. It has 1 scale on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 272,
+ 349,
+ 281,
+ 358
+ ],
+ "score": 0.75,
+ "content": "T",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 281,
+ 347,
+ 343,
+ 361
+ ],
+ "score": 1.0,
+ "content": "and 4 scales on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 349,
+ 354,
+ 358
+ ],
+ "score": 0.74,
+ "content": "H",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 354,
+ 347,
+ 459,
+ 361
+ ],
+ "score": 1.0,
+ "content": ". It is convolved with filter",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 459,
+ 349,
+ 504,
+ 359
+ ],
+ "score": 0.9,
+ "content": "\\kappa = w \\times \\Psi",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 360,
+ 505,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 360,
+ 505,
+ 371
+ ],
+ "score": 1.0,
+ "content": "without scale interaction, which produces the output with 2 channels and 4 scales as well. Here we",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 370,
+ 497,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 370,
+ 497,
+ 383
+ ],
+ "score": 1.0,
+ "content": "represent only channels of the signals and the filter. Spatial components are hidden for simplicity.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 8.0
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 408,
+ 211,
+ 421
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 213,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 213,
+ 423
+ ],
+ "score": 1.0,
+ "content": "5 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 434,
+ 505,
+ 566
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 434,
+ 506,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 434,
+ 506,
+ 448
+ ],
+ "score": 1.0,
+ "content": "Various works on group-equivariant convolutional networks have been published recently. These",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 446,
+ 505,
+ 458
+ ],
+ "score": 1.0,
+ "content": "works have considered roto-translation groups in 2D Cohen & Welling (2016a); Hoogeboom et al.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 457,
+ 505,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 457,
+ 505,
+ 468
+ ],
+ "score": 1.0,
+ "content": "(2018); Worrall et al. (2017); Weiler & Cesa (2019) and 3D Worrall & Brostow (2018); Kondor",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 466,
+ 506,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 506,
+ 480
+ ],
+ "score": 1.0,
+ "content": "(2018); Thomas et al. (2018) and rotation equivariant networks in 3D Cohen et al. (2017); Esteves",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 478,
+ 505,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 478,
+ 505,
+ 491
+ ],
+ "score": 1.0,
+ "content": "et al. (2018); Cohen et al. (2019). In Freeman & Adelson (1991) authors describe the algorithm",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 104,
+ 488,
+ 506,
+ 503
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 488,
+ 506,
+ 503
+ ],
+ "score": 1.0,
+ "content": "for designing steerable filters for rotations. Rotation steerable filters are used in Cohen & Welling",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 513
+ ],
+ "score": 1.0,
+ "content": "(2016b); Weiler et al. (2018a;b) for building equivariant networks. In Jacobsen et al. (2017) the",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 510,
+ 506,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 510,
+ 369,
+ 525
+ ],
+ "score": 1.0,
+ "content": "authors build convolutional blocks locally equivariant to arbitrary",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 369,
+ 512,
+ 376,
+ 521
+ ],
+ "score": 0.81,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 376,
+ 510,
+ 506,
+ 525
+ ],
+ "score": 1.0,
+ "content": "-parameter Lie group by using a",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 535
+ ],
+ "score": 1.0,
+ "content": "steerable basis. And in Murugan et al. the authors discuss the approach for learning steerable filters",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 104,
+ 532,
+ 506,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 532,
+ 506,
+ 546
+ ],
+ "score": 1.0,
+ "content": "from data. To date, the majority of papers on group equivariant networks have considered rotations",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 104,
+ 543,
+ 506,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 543,
+ 506,
+ 558
+ ],
+ "score": 1.0,
+ "content": "in 2D and 3D, but have not payed attention to scale symmetry. As we have argued above, it is a",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 225,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 225,
+ 568
+ ],
+ "score": 1.0,
+ "content": "fundamentally different case.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 19.5,
+ "bbox_fs": [
+ 104,
+ 434,
+ 506,
+ 568
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 572,
+ 505,
+ 649
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 572,
+ 505,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 572,
+ 505,
+ 585
+ ],
+ "score": 1.0,
+ "content": "Many papers and even conferences have been dedicated to image scale-space — a concept where",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 583,
+ 505,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 583,
+ 505,
+ 595
+ ],
+ "score": 1.0,
+ "content": "the image is analyzed together with all its downscaled versions. Initially introduced in Iijima (1959)",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 607
+ ],
+ "score": 1.0,
+ "content": "and later developed by Witkin (1987); Perona & Malik (1990); Lindeberg (2013) scale space relies",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "on the scale symmetry of images. The differential structure of the image Koenderink (1984) allows",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "one to make a connection between image formation mechanisms and the space of solutions of the 2-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 641
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 641
+ ],
+ "score": 1.0,
+ "content": "dimensional heat equation, which significantly improved the image analysis models in the pre-deep",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 159,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 159,
+ 652
+ ],
+ "score": 1.0,
+ "content": "learning era.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 29,
+ "bbox_fs": [
+ 105,
+ 572,
+ 506,
+ 652
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 654,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 654,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 654,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "One of the first works on scale equivariance and local scale invariance in the framework of CNNs was",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 505,
+ 677
+ ],
+ "score": 1.0,
+ "content": "proposed by Xu et al. (2014) named SiCNN. The authors describe the model with siamese CNNs,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "where the filters of each instance are rescaled using interpolation techniques. This is the simplest",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "case of equivariance where no interaction between different scales is done in intermediate layers. In",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "SI-ConvNet by Kanazawa et al. (2014) the original network is modified such that, in each layer, the",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "input is first rescaled, then convolved and rescaled back to the original size. Finally, the response",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 720,
+ 505,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 720,
+ 505,
+ 733
+ ],
+ "score": 1.0,
+ "content": "with the maximum values is chosen between the scales. Thus, the model is locally scale-invariant. In",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 36,
+ "bbox_fs": [
+ 105,
+ 654,
+ 506,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 135,
+ 80,
+ 476,
+ 175
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 135,
+ 80,
+ 476,
+ 175
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 135,
+ 80,
+ 476,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 135,
+ 80,
+ 476,
+ 175
+ ],
+ "score": 0.982,
+ "html": "| Method | Equivariance | Admissible Scales | Approach | Interscale |
| SiCNN | | Grid | Filter Rescaling | |
| SI-ConvNet | | Grid | Input Rescaling | |
| SEVF | | Grid | Input Rescaling | |
| DSS | <x</ | Integer | Filter Dilation | xx<> |
| SS-CNN | X | Any | Steerable Filters | X |
| SESN, Ours | √ | Any | Steerable Filters | √ |
",
+ "type": "table",
+ "image_path": "034542743359fb7a7b899fde9ba50daf8976a4bd6ccd05c94a4e0a5a4ca09f77.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 135,
+ 80,
+ 476,
+ 111.66666666666667
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 135,
+ 111.66666666666667,
+ 476,
+ 143.33333333333334
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 135,
+ 143.33333333333334,
+ 476,
+ 175.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 107,
+ 187,
+ 505,
+ 232
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 506,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 259,
+ 201
+ ],
+ "score": 1.0,
+ "content": "Table 1: Comparing SESN to SiCNN",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 188,
+ 274,
+ 198
+ ],
+ "score": 0.33,
+ "content": "\\mathrm { X u }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 187,
+ 506,
+ 201
+ ],
+ "score": 1.0,
+ "content": "et al. (2014), SI-ConvNet Kanazawa et al. (2014), SEVF",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 199,
+ 505,
+ 211
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 505,
+ 211
+ ],
+ "score": 1.0,
+ "content": "Marcos et al. (2018), DSS Worrall & Welling (2019) and SS-CNN Ghosh & Gupta (2019). “In-",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 210,
+ 505,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 210,
+ 505,
+ 222
+ ],
+ "score": 1.0,
+ "content": "terscale” refers to the ability of capturing interscale interactions with kernels of non-unitary scale",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 221,
+ 504,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 221,
+ 504,
+ 233
+ ],
+ "score": 1.0,
+ "content": "extent. “Grid” stands for the scales which generate images which lie exactly on the initial pixel grid.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5
+ }
+ ],
+ "index": 2.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 259,
+ 505,
+ 359
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 505,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 505,
+ 272
+ ],
+ "score": 1.0,
+ "content": "Marcos et al. (2018), in the SEVF model, the input of the layers is rescaled and convolved multiple",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 272,
+ 505,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 272,
+ 505,
+ 284
+ ],
+ "score": 1.0,
+ "content": "times to form vector features instead of scalar ones. The length of the vector in each position is the",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 283,
+ 505,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 283,
+ 505,
+ 294
+ ],
+ "score": 1.0,
+ "content": "maximum magnitude of the convolution, while the direction of the angle encodes the scale of the",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 293,
+ 506,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 293,
+ 506,
+ 306
+ ],
+ "score": 1.0,
+ "content": "image which gave this response. These scale-equivariant networks rely on image rescaling which is",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 304,
+ 505,
+ 318
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 304,
+ 505,
+ 318
+ ],
+ "score": 1.0,
+ "content": "quite slow. Worrall & Welling (2019) (DSS) generalize the concept of scale-space to deep networks.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 314,
+ 506,
+ 328
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 314,
+ 506,
+ 328
+ ],
+ "score": 1.0,
+ "content": "They use filter dilation to analyze the images on different scales. While this approach is as fast as",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 326,
+ 505,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 326,
+ 381,
+ 339
+ ],
+ "score": 1.0,
+ "content": "the standard CNN, it is restricted only to integer downscale factors",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 381,
+ 326,
+ 424,
+ 338
+ ],
+ "score": 0.45,
+ "content": "2 , 4 , 8 \\dots",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 424,
+ 326,
+ 505,
+ 339
+ ],
+ "score": 1.0,
+ "content": "In Ghosh & Gupta",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 336,
+ 505,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 336,
+ 505,
+ 350
+ ],
+ "score": 1.0,
+ "content": "(2019), while discussing SS-CNN the authors use scale-steerable filters to deal with scale changes.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 347,
+ 427,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 347,
+ 427,
+ 361
+ ],
+ "score": 1.0,
+ "content": "The paper does not discuss equivariance, which is an important aspect for scale.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 365,
+ 505,
+ 409
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 364,
+ 505,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 364,
+ 505,
+ 377
+ ],
+ "score": 1.0,
+ "content": "We summarize the information about these models in Table 1. In contrast to other scale-equivariant",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 376,
+ 505,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 376,
+ 505,
+ 387
+ ],
+ "score": 1.0,
+ "content": "models, SESN uses steerable filters which allows for fast scale-convolution with no limitation of",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 386,
+ 505,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 386,
+ 505,
+ 398
+ ],
+ "score": 1.0,
+ "content": "flexibility. With the framework of Scale-Equivariant Convolutional Networks we are free to build",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 398,
+ 333,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 398,
+ 333,
+ 410
+ ],
+ "score": 1.0,
+ "content": "both equivariant and invariant models of different kinds.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 17.5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 113,
+ 478,
+ 497,
+ 575
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 113,
+ 478,
+ 497,
+ 575
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 113,
+ 478,
+ 497,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 478,
+ 497,
+ 575
+ ],
+ "score": 0.964,
+ "type": "image",
+ "image_path": "9d5081dc572954800739488564f5c9304a98127ff3d55fe8a82463473ff120dc.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 113,
+ 478,
+ 497,
+ 510.3333333333333
+ ],
+ "spans": [],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 113,
+ 510.3333333333333,
+ 497,
+ 542.6666666666666
+ ],
+ "spans": [],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 113,
+ 542.6666666666666,
+ 497,
+ 575.0
+ ],
+ "spans": [],
+ "index": 22
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 594,
+ 505,
+ 627
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 226,
+ 607
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Equivariance error",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 227,
+ 595,
+ 236,
+ 604
+ ],
+ "score": 0.79,
+ "content": "\\Delta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 237,
+ 594,
+ 506,
+ 607
+ ],
+ "score": 1.0,
+ "content": "as a function of the number of layers (left), downscaling applied",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "to the input image (middle), and as a function of number of scales in interscale interactions (right).",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 256,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 256,
+ 628
+ ],
+ "score": 1.0,
+ "content": "The bars indicate standard deviation.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 22.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 649,
+ 505,
+ 693
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 505,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 505,
+ 662
+ ],
+ "score": 1.0,
+ "content": "In this section we conduct the experiments and compare various methods for working with scale",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 660,
+ 505,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 505,
+ 672
+ ],
+ "score": 1.0,
+ "content": "variations in input data. Alongside SESN, we test local scale invariant SI-ConvNet and SS-CNN,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 505,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 505,
+ 683
+ ],
+ "score": 1.0,
+ "content": "scale equivariant SiCNN, SEVF and DSS. For SEVF, DSS and SS-CNN we use the code provided",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 682,
+ 394,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 682,
+ 394,
+ 694
+ ],
+ "score": 1.0,
+ "content": "by authors, while for others we reimplement the main buildings blocks.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 27.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 699,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 504,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 504,
+ 711
+ ],
+ "score": 1.0,
+ "content": "We provide additional experimental results on time performance of all these methods in Appendix B.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "Due to the algorithm proposed in Section 4 SESN allows for training several times faster than other",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 267,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 267,
+ 733
+ ],
+ "score": 1.0,
+ "content": "methods which rely on image rescaling.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "page_idx": 5,
+ "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 2020",
+ "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"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 433,
+ 200,
+ 444
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 431,
+ 202,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 431,
+ 202,
+ 447
+ ],
+ "score": 1.0,
+ "content": "6 EXPERIMENTS",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 135,
+ 80,
+ 476,
+ 175
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 135,
+ 80,
+ 476,
+ 175
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 135,
+ 80,
+ 476,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 135,
+ 80,
+ 476,
+ 175
+ ],
+ "score": 0.982,
+ "html": "| Method | Equivariance | Admissible Scales | Approach | Interscale |
| SiCNN | | Grid | Filter Rescaling | |
| SI-ConvNet | | Grid | Input Rescaling | |
| SEVF | | Grid | Input Rescaling | |
| DSS | <x</ | Integer | Filter Dilation | xx<> |
| SS-CNN | X | Any | Steerable Filters | X |
| SESN, Ours | √ | Any | Steerable Filters | √ |
",
+ "type": "table",
+ "image_path": "034542743359fb7a7b899fde9ba50daf8976a4bd6ccd05c94a4e0a5a4ca09f77.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 135,
+ 80,
+ 476,
+ 111.66666666666667
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 135,
+ 111.66666666666667,
+ 476,
+ 143.33333333333334
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 135,
+ 143.33333333333334,
+ 476,
+ 175.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 107,
+ 187,
+ 505,
+ 232
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 506,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 259,
+ 201
+ ],
+ "score": 1.0,
+ "content": "Table 1: Comparing SESN to SiCNN",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 259,
+ 188,
+ 274,
+ 198
+ ],
+ "score": 0.33,
+ "content": "\\mathrm { X u }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 187,
+ 506,
+ 201
+ ],
+ "score": 1.0,
+ "content": "et al. (2014), SI-ConvNet Kanazawa et al. (2014), SEVF",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 199,
+ 505,
+ 211
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 505,
+ 211
+ ],
+ "score": 1.0,
+ "content": "Marcos et al. (2018), DSS Worrall & Welling (2019) and SS-CNN Ghosh & Gupta (2019). “In-",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 210,
+ 505,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 210,
+ 505,
+ 222
+ ],
+ "score": 1.0,
+ "content": "terscale” refers to the ability of capturing interscale interactions with kernels of non-unitary scale",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 221,
+ 504,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 221,
+ 504,
+ 233
+ ],
+ "score": 1.0,
+ "content": "extent. “Grid” stands for the scales which generate images which lie exactly on the initial pixel grid.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5
+ }
+ ],
+ "index": 2.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 259,
+ 505,
+ 359
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 505,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 259,
+ 505,
+ 272
+ ],
+ "score": 1.0,
+ "content": "Marcos et al. (2018), in the SEVF model, the input of the layers is rescaled and convolved multiple",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 272,
+ 505,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 272,
+ 505,
+ 284
+ ],
+ "score": 1.0,
+ "content": "times to form vector features instead of scalar ones. The length of the vector in each position is the",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 283,
+ 505,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 283,
+ 505,
+ 294
+ ],
+ "score": 1.0,
+ "content": "maximum magnitude of the convolution, while the direction of the angle encodes the scale of the",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 293,
+ 506,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 293,
+ 506,
+ 306
+ ],
+ "score": 1.0,
+ "content": "image which gave this response. These scale-equivariant networks rely on image rescaling which is",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 304,
+ 505,
+ 318
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 304,
+ 505,
+ 318
+ ],
+ "score": 1.0,
+ "content": "quite slow. Worrall & Welling (2019) (DSS) generalize the concept of scale-space to deep networks.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 314,
+ 506,
+ 328
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 314,
+ 506,
+ 328
+ ],
+ "score": 1.0,
+ "content": "They use filter dilation to analyze the images on different scales. While this approach is as fast as",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 326,
+ 505,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 326,
+ 381,
+ 339
+ ],
+ "score": 1.0,
+ "content": "the standard CNN, it is restricted only to integer downscale factors",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 381,
+ 326,
+ 424,
+ 338
+ ],
+ "score": 0.45,
+ "content": "2 , 4 , 8 \\dots",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 424,
+ 326,
+ 505,
+ 339
+ ],
+ "score": 1.0,
+ "content": "In Ghosh & Gupta",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 336,
+ 505,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 336,
+ 505,
+ 350
+ ],
+ "score": 1.0,
+ "content": "(2019), while discussing SS-CNN the authors use scale-steerable filters to deal with scale changes.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 347,
+ 427,
+ 361
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 347,
+ 427,
+ 361
+ ],
+ "score": 1.0,
+ "content": "The paper does not discuss equivariance, which is an important aspect for scale.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 11,
+ "bbox_fs": [
+ 105,
+ 259,
+ 506,
+ 361
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 365,
+ 505,
+ 409
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 364,
+ 505,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 364,
+ 505,
+ 377
+ ],
+ "score": 1.0,
+ "content": "We summarize the information about these models in Table 1. In contrast to other scale-equivariant",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 376,
+ 505,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 376,
+ 505,
+ 387
+ ],
+ "score": 1.0,
+ "content": "models, SESN uses steerable filters which allows for fast scale-convolution with no limitation of",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 386,
+ 505,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 386,
+ 505,
+ 398
+ ],
+ "score": 1.0,
+ "content": "flexibility. With the framework of Scale-Equivariant Convolutional Networks we are free to build",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 398,
+ 333,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 398,
+ 333,
+ 410
+ ],
+ "score": 1.0,
+ "content": "both equivariant and invariant models of different kinds.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 17.5,
+ "bbox_fs": [
+ 105,
+ 364,
+ 505,
+ 410
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 113,
+ 478,
+ 497,
+ 575
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 113,
+ 478,
+ 497,
+ 575
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 113,
+ 478,
+ 497,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 478,
+ 497,
+ 575
+ ],
+ "score": 0.964,
+ "type": "image",
+ "image_path": "9d5081dc572954800739488564f5c9304a98127ff3d55fe8a82463473ff120dc.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 21,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 113,
+ 478,
+ 497,
+ 510.3333333333333
+ ],
+ "spans": [],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 113,
+ 510.3333333333333,
+ 497,
+ 542.6666666666666
+ ],
+ "spans": [],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 113,
+ 542.6666666666666,
+ 497,
+ 575.0
+ ],
+ "spans": [],
+ "index": 22
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 594,
+ 505,
+ 627
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 506,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 226,
+ 607
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Equivariance error",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 227,
+ 595,
+ 236,
+ 604
+ ],
+ "score": 0.79,
+ "content": "\\Delta",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 237,
+ 594,
+ 506,
+ 607
+ ],
+ "score": 1.0,
+ "content": "as a function of the number of layers (left), downscaling applied",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "to the input image (middle), and as a function of number of scales in interscale interactions (right).",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 256,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 256,
+ 628
+ ],
+ "score": 1.0,
+ "content": "The bars indicate standard deviation.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 22.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 649,
+ 505,
+ 693
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 505,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 505,
+ 662
+ ],
+ "score": 1.0,
+ "content": "In this section we conduct the experiments and compare various methods for working with scale",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 660,
+ 505,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 505,
+ 672
+ ],
+ "score": 1.0,
+ "content": "variations in input data. Alongside SESN, we test local scale invariant SI-ConvNet and SS-CNN,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 505,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 505,
+ 683
+ ],
+ "score": 1.0,
+ "content": "scale equivariant SiCNN, SEVF and DSS. For SEVF, DSS and SS-CNN we use the code provided",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 682,
+ 394,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 682,
+ 394,
+ 694
+ ],
+ "score": 1.0,
+ "content": "by authors, while for others we reimplement the main buildings blocks.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 27.5,
+ "bbox_fs": [
+ 105,
+ 649,
+ 505,
+ 694
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 699,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 504,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 504,
+ 711
+ ],
+ "score": 1.0,
+ "content": "We provide additional experimental results on time performance of all these methods in Appendix B.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "Due to the algorithm proposed in Section 4 SESN allows for training several times faster than other",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 267,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 267,
+ 733
+ ],
+ "score": 1.0,
+ "content": "methods which rely on image rescaling.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 31,
+ "bbox_fs": [
+ 106,
+ 699,
+ 505,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 114,
+ 80,
+ 496,
+ 219
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 114,
+ 80,
+ 496,
+ 219
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 114,
+ 80,
+ 496,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 80,
+ 496,
+ 219
+ ],
+ "score": 0.985,
+ "html": "| Method | (28×28) | (28 × 28) + | (56 × 56) | (56 × 56) + | # Params |
| CNN | 2.56 ± 0.04 | 1.96 ± 0.07 | 2.02 ± 0.07 | 1.60 ± 0.09 | 495K |
| SiCNN | 2.40 ± 0.03 | 1.86 ± 0.10 | 2.02 ± 0.14 | 1.59 ± 0.03 | 497K |
| SI-ConvNet | 2.40± 0.12 | 1.94 ± 0.07 | 1.82 ± 0.11 | 1.59 ± 0.10 | 495K |
| SEVF Scalar | 2.30 ± 0.06 | 1.96 ± 0.07 | 1.87 ± 0.09 | 1.62 ± 0.07 | 494 K |
| SEVF Vector | 2.63 ± 0.09 | 2.23 ± 0.09 | 2.12 ± 0.13 | 1.81 ± 0.09 | 475 K |
| DSS Scalar | 2.53 ± 0.10 | 2.04± 0.08 | 1.92 ± 0.08 | 1.57 ± 0.08 | 494 K |
| DSS Vector | 2.58 ± 0.11 | 1.95 ± 0.07 | 1.97 ± 0.08 | 1.57 ± 0.09 | 494 K |
| SS-CNN | 2.32 ± 0.15 | 2.10 ±0.15 | 1.84 ± 0.10 | 1.76 ± 0.07 | 494 K |
| SESN Scalar | 2.10 ±0.10 | 1.79 ± 0.09 | 1.74 ± 0.09 | 1.50 ± 0.07 | 495K |
| SESN Vector | 2.08 ±0.09 | 1.76 ± 0.08 | 1.68 ± 0.06 | 1.42 ± 0.07 | 495 K |
",
+ "type": "table",
+ "image_path": "f969ef34917f25cff64cb98c019f110ea15b35f75cc339524e69d9b6629f4ddb.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 114,
+ 80,
+ 496,
+ 126.33333333333334
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 114,
+ 126.33333333333334,
+ 496,
+ 172.66666666666669
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 114,
+ 172.66666666666669,
+ 496,
+ 219.00000000000003
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 232,
+ 505,
+ 276
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 231,
+ 505,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 231,
+ 505,
+ 244
+ ],
+ "score": 1.0,
+ "content": "Table 2: Classification error of different methods on MNIST-scale dataset, lower is better. In ex-",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 104,
+ 243,
+ 506,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 243,
+ 258,
+ 255
+ ],
+ "score": 1.0,
+ "content": "periment we use image resolution of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 258,
+ 243,
+ 293,
+ 253
+ ],
+ "score": 0.91,
+ "content": "2 8 \\times 2 8",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 293,
+ 243,
+ 312,
+ 255
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 312,
+ 243,
+ 346,
+ 254
+ ],
+ "score": 0.9,
+ "content": "5 6 \\times 5 6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 346,
+ 243,
+ 506,
+ 255
+ ],
+ "score": 1.0,
+ "content": ". We test both the regime without data",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 254,
+ 505,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 254,
+ 421,
+ 266
+ ],
+ "score": 1.0,
+ "content": "augmentation, and the regime with scaling data augmentation, denoted with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 421,
+ 254,
+ 438,
+ 265
+ ],
+ "score": 0.56,
+ "content": "\" + \"",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 438,
+ 254,
+ 505,
+ 266
+ ],
+ "score": 1.0,
+ "content": ". All results are",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 265,
+ 498,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 265,
+ 176,
+ 276
+ ],
+ "score": 1.0,
+ "content": "reported as mean",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 176,
+ 266,
+ 187,
+ 275
+ ],
+ "score": 0.74,
+ "content": "\\pm",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 187,
+ 265,
+ 498,
+ 276
+ ],
+ "score": 1.0,
+ "content": "std over 6 different fixed realizations of the dataset. The best results are bold.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 301,
+ 229,
+ 312
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 300,
+ 231,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 300,
+ 231,
+ 313
+ ],
+ "score": 1.0,
+ "content": "6.1 EQUIVARIANCE ERROR",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 323,
+ 505,
+ 389
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 324,
+ 505,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 324,
+ 505,
+ 335
+ ],
+ "score": 1.0,
+ "content": "We have presented scale-convolution which is equivariant to scale transformation and translation for",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 334,
+ 505,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 505,
+ 347
+ ],
+ "score": 1.0,
+ "content": "continuous signals. While translation equivariance holds true even for discretized signals and filters,",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 345,
+ 505,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 505,
+ 357
+ ],
+ "score": 1.0,
+ "content": "scale equivariance may not be exact. Therefore, before starting any experiments, we check to which",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 356,
+ 505,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 356,
+ 505,
+ 369
+ ],
+ "score": 1.0,
+ "content": "degree the predicted properties of scale-convolution hold true. We do so by measuring the difference",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 365,
+ 506,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 366,
+ 276,
+ 379
+ ],
+ "score": 0.93,
+ "content": "\\dot { \\Delta } = \\| [ L _ { s } \\mathbf { \\hat { \\Phi } } ( f ) - \\Phi \\mathbf { \\hat { L } } _ { s } ( f ) \\| _ { 2 } ^ { 2 } / \\| L _ { s } \\Phi ( f ) \\| _ { 2 } ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 277,
+ 365,
+ 308,
+ 380
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 308,
+ 367,
+ 317,
+ 377
+ ],
+ "score": 0.84,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 317,
+ 365,
+ 506,
+ 380
+ ],
+ "score": 1.0,
+ "content": "is scale-convolution with randomly initialized",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 378,
+ 143,
+ 391
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 378,
+ 143,
+ 391
+ ],
+ "score": 1.0,
+ "content": "weights.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 10.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 394,
+ 505,
+ 526
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 395,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 395,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "In case of perfect equivariance the difference is equal to zero. We calculate the error on randomly",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 419
+ ],
+ "score": 1.0,
+ "content": "sampled images from the STL-10 dataset Coates et al. (2011). The results are represented in Fig-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 417,
+ 505,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 417,
+ 505,
+ 429
+ ],
+ "score": 1.0,
+ "content": "ure 2. The networks on the left and on the middle plots do not have interscale interactions. The",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 505,
+ 440
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 420,
+ 440
+ ],
+ "score": 1.0,
+ "content": "networks on the middle and on the right plots consist of just one layer. We use",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 420,
+ 428,
+ 478,
+ 439
+ ],
+ "score": 0.79,
+ "content": "N _ { S } = 5 , 1 3 , 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 478,
+ 428,
+ 505,
+ 440
+ ],
+ "score": 1.0,
+ "content": "scales",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 439,
+ 505,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 439,
+ 505,
+ 451
+ ],
+ "score": 1.0,
+ "content": "for the networks on the left, the middle, and the right plots respectively. While discretization intro-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 450,
+ 505,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 450,
+ 360,
+ 462
+ ],
+ "score": 1.0,
+ "content": "duces some error, it stays very low, and is not much higher than",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 361,
+ 450,
+ 375,
+ 460
+ ],
+ "score": 0.85,
+ "content": "6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 376,
+ 450,
+ 505,
+ 462
+ ],
+ "score": 1.0,
+ "content": "for the networks with 50 layers.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 473
+ ],
+ "score": 1.0,
+ "content": "The difference, however, increases if the input image is downscaled more than 16 times. There-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 484
+ ],
+ "score": 1.0,
+ "content": "fore, we are free to use deep networks. However, we should pay extra attention to extreme cases",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 495
+ ],
+ "score": 1.0,
+ "content": "where scale changes are of very big magnitude. These are quite rare but still appear in practice.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 493,
+ 505,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 493,
+ 505,
+ 505
+ ],
+ "score": 1.0,
+ "content": "Finally, we see that using SESN with interscale interaction introduces extra equivariance error due",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 505,
+ 505,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 505,
+ 185,
+ 516
+ ],
+ "score": 1.0,
+ "content": "to the truncation of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 185,
+ 505,
+ 192,
+ 514
+ ],
+ "score": 0.76,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 193,
+ 505,
+ 505,
+ 516
+ ],
+ "score": 1.0,
+ "content": ". We will build the networks with either no scale interaction or interaction of 2",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 516,
+ 136,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 516,
+ 136,
+ 527
+ ],
+ "score": 1.0,
+ "content": "scales.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 19.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 543,
+ 196,
+ 555
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 542,
+ 198,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 542,
+ 198,
+ 558
+ ],
+ "score": 1.0,
+ "content": "6.2 MNIST-SCALE",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 565,
+ 505,
+ 632
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 566,
+ 505,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 566,
+ 505,
+ 578
+ ],
+ "score": 1.0,
+ "content": "Following Kanazawa et al. (2014); Marcos et al. (2018); Ghosh & Gupta (2019) we conduct experi-",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 505,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 505,
+ 589
+ ],
+ "score": 1.0,
+ "content": "ments on the MNIST-scale dataset. We rescale the images of the MNIST dataset LeCun et al. (1998)",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 588,
+ 505,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 588,
+ 117,
+ 600
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 117,
+ 588,
+ 157,
+ 599
+ ],
+ "score": 0.78,
+ "content": "0 . 3 - 1 . 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 158,
+ 588,
+ 505,
+ 600
+ ],
+ "score": 1.0,
+ "content": "of the original size and pad them with zeros to retain the initial resolution. The scal-",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 599,
+ 505,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 599,
+ 505,
+ 612
+ ],
+ "score": 1.0,
+ "content": "ing factors are sampled uniformly and independently for each image. The obtained dataset is then",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 622
+ ],
+ "score": 1.0,
+ "content": "split into 10,000 for training, 2,000 for evaluation and 50,000 for testing. We generate 6 different",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 621,
+ 287,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 287,
+ 633
+ ],
+ "score": 1.0,
+ "content": "realizations and fix them for all experiments.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 637,
+ 505,
+ 693
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 638,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 638,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "As a baseline model we use the model described in Ghosh & Gupta (2019), which currently holds",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 648,
+ 505,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 505,
+ 662
+ ],
+ "score": 1.0,
+ "content": "the state-of-the-art result on this dataset. It consists of 3 convolutional and 2 fully-connected layers.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 659,
+ 505,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 659,
+ 224,
+ 673
+ ],
+ "score": 1.0,
+ "content": "Each layer has filters of size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 225,
+ 660,
+ 249,
+ 671
+ ],
+ "score": 0.88,
+ "content": "7 \\times 7",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 249,
+ 659,
+ 505,
+ 673
+ ],
+ "score": 1.0,
+ "content": ". We keep the number of trainable parameters almost the same",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 671,
+ 505,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 671,
+ 505,
+ 683
+ ],
+ "score": 1.0,
+ "content": "for all tested methods. This is achieved by varying the number of channels. For scale equivariant",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 681,
+ 384,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 681,
+ 384,
+ 694
+ ],
+ "score": 1.0,
+ "content": "models we add scale projection at the end of the convolutional block.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 699,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "For SiCNN, DSS, SEVF and our model, we additionally train counterparts where after each con-",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 709,
+ 506,
+ 721
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 506,
+ 721
+ ],
+ "score": 1.0,
+ "content": "volution, an extra projection layer is inserted. Projection layers transform vector features in each",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 733
+ ],
+ "score": 1.0,
+ "content": "spatial position of each channel into scalar ones. All of the layers have now scalar inputs instead of",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "page_idx": 6,
+ "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 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "7",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 114,
+ 80,
+ 496,
+ 219
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 114,
+ 80,
+ 496,
+ 219
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 114,
+ 80,
+ 496,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 80,
+ 496,
+ 219
+ ],
+ "score": 0.985,
+ "html": "| Method | (28×28) | (28 × 28) + | (56 × 56) | (56 × 56) + | # Params |
| CNN | 2.56 ± 0.04 | 1.96 ± 0.07 | 2.02 ± 0.07 | 1.60 ± 0.09 | 495K |
| SiCNN | 2.40 ± 0.03 | 1.86 ± 0.10 | 2.02 ± 0.14 | 1.59 ± 0.03 | 497K |
| SI-ConvNet | 2.40± 0.12 | 1.94 ± 0.07 | 1.82 ± 0.11 | 1.59 ± 0.10 | 495K |
| SEVF Scalar | 2.30 ± 0.06 | 1.96 ± 0.07 | 1.87 ± 0.09 | 1.62 ± 0.07 | 494 K |
| SEVF Vector | 2.63 ± 0.09 | 2.23 ± 0.09 | 2.12 ± 0.13 | 1.81 ± 0.09 | 475 K |
| DSS Scalar | 2.53 ± 0.10 | 2.04± 0.08 | 1.92 ± 0.08 | 1.57 ± 0.08 | 494 K |
| DSS Vector | 2.58 ± 0.11 | 1.95 ± 0.07 | 1.97 ± 0.08 | 1.57 ± 0.09 | 494 K |
| SS-CNN | 2.32 ± 0.15 | 2.10 ±0.15 | 1.84 ± 0.10 | 1.76 ± 0.07 | 494 K |
| SESN Scalar | 2.10 ±0.10 | 1.79 ± 0.09 | 1.74 ± 0.09 | 1.50 ± 0.07 | 495K |
| SESN Vector | 2.08 ±0.09 | 1.76 ± 0.08 | 1.68 ± 0.06 | 1.42 ± 0.07 | 495 K |
",
+ "type": "table",
+ "image_path": "f969ef34917f25cff64cb98c019f110ea15b35f75cc339524e69d9b6629f4ddb.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 114,
+ 80,
+ 496,
+ 126.33333333333334
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 114,
+ 126.33333333333334,
+ 496,
+ 172.66666666666669
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 114,
+ 172.66666666666669,
+ 496,
+ 219.00000000000003
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 232,
+ 505,
+ 276
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 231,
+ 505,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 231,
+ 505,
+ 244
+ ],
+ "score": 1.0,
+ "content": "Table 2: Classification error of different methods on MNIST-scale dataset, lower is better. In ex-",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 104,
+ 243,
+ 506,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 243,
+ 258,
+ 255
+ ],
+ "score": 1.0,
+ "content": "periment we use image resolution of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 258,
+ 243,
+ 293,
+ 253
+ ],
+ "score": 0.91,
+ "content": "2 8 \\times 2 8",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 293,
+ 243,
+ 312,
+ 255
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 312,
+ 243,
+ 346,
+ 254
+ ],
+ "score": 0.9,
+ "content": "5 6 \\times 5 6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 346,
+ 243,
+ 506,
+ 255
+ ],
+ "score": 1.0,
+ "content": ". We test both the regime without data",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 254,
+ 505,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 254,
+ 421,
+ 266
+ ],
+ "score": 1.0,
+ "content": "augmentation, and the regime with scaling data augmentation, denoted with",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 421,
+ 254,
+ 438,
+ 265
+ ],
+ "score": 0.56,
+ "content": "\" + \"",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 438,
+ 254,
+ 505,
+ 266
+ ],
+ "score": 1.0,
+ "content": ". All results are",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 265,
+ 498,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 265,
+ 176,
+ 276
+ ],
+ "score": 1.0,
+ "content": "reported as mean",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 176,
+ 266,
+ 187,
+ 275
+ ],
+ "score": 0.74,
+ "content": "\\pm",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 187,
+ 265,
+ 498,
+ 276
+ ],
+ "score": 1.0,
+ "content": "std over 6 different fixed realizations of the dataset. The best results are bold.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5,
+ "bbox_fs": [
+ 104,
+ 231,
+ 506,
+ 276
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 301,
+ 229,
+ 312
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 300,
+ 231,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 300,
+ 231,
+ 313
+ ],
+ "score": 1.0,
+ "content": "6.1 EQUIVARIANCE ERROR",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 323,
+ 505,
+ 389
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 324,
+ 505,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 324,
+ 505,
+ 335
+ ],
+ "score": 1.0,
+ "content": "We have presented scale-convolution which is equivariant to scale transformation and translation for",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 334,
+ 505,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 334,
+ 505,
+ 347
+ ],
+ "score": 1.0,
+ "content": "continuous signals. While translation equivariance holds true even for discretized signals and filters,",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 345,
+ 505,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 505,
+ 357
+ ],
+ "score": 1.0,
+ "content": "scale equivariance may not be exact. Therefore, before starting any experiments, we check to which",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 356,
+ 505,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 356,
+ 505,
+ 369
+ ],
+ "score": 1.0,
+ "content": "degree the predicted properties of scale-convolution hold true. We do so by measuring the difference",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 365,
+ 506,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 366,
+ 276,
+ 379
+ ],
+ "score": 0.93,
+ "content": "\\dot { \\Delta } = \\| [ L _ { s } \\mathbf { \\hat { \\Phi } } ( f ) - \\Phi \\mathbf { \\hat { L } } _ { s } ( f ) \\| _ { 2 } ^ { 2 } / \\| L _ { s } \\Phi ( f ) \\| _ { 2 } ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 277,
+ 365,
+ 308,
+ 380
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 308,
+ 367,
+ 317,
+ 377
+ ],
+ "score": 0.84,
+ "content": "\\Phi",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 317,
+ 365,
+ 506,
+ 380
+ ],
+ "score": 1.0,
+ "content": "is scale-convolution with randomly initialized",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 378,
+ 143,
+ 391
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 378,
+ 143,
+ 391
+ ],
+ "score": 1.0,
+ "content": "weights.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 10.5,
+ "bbox_fs": [
+ 105,
+ 324,
+ 506,
+ 391
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 394,
+ 505,
+ 526
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 395,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 395,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "In case of perfect equivariance the difference is equal to zero. We calculate the error on randomly",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 419
+ ],
+ "score": 1.0,
+ "content": "sampled images from the STL-10 dataset Coates et al. (2011). The results are represented in Fig-",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 417,
+ 505,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 417,
+ 505,
+ 429
+ ],
+ "score": 1.0,
+ "content": "ure 2. The networks on the left and on the middle plots do not have interscale interactions. The",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 428,
+ 505,
+ 440
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 428,
+ 420,
+ 440
+ ],
+ "score": 1.0,
+ "content": "networks on the middle and on the right plots consist of just one layer. We use",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 420,
+ 428,
+ 478,
+ 439
+ ],
+ "score": 0.79,
+ "content": "N _ { S } = 5 , 1 3 , 5",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 478,
+ 428,
+ 505,
+ 440
+ ],
+ "score": 1.0,
+ "content": "scales",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 439,
+ 505,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 439,
+ 505,
+ 451
+ ],
+ "score": 1.0,
+ "content": "for the networks on the left, the middle, and the right plots respectively. While discretization intro-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 450,
+ 505,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 450,
+ 360,
+ 462
+ ],
+ "score": 1.0,
+ "content": "duces some error, it stays very low, and is not much higher than",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 361,
+ 450,
+ 375,
+ 460
+ ],
+ "score": 0.85,
+ "content": "6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 376,
+ 450,
+ 505,
+ 462
+ ],
+ "score": 1.0,
+ "content": "for the networks with 50 layers.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 473
+ ],
+ "score": 1.0,
+ "content": "The difference, however, increases if the input image is downscaled more than 16 times. There-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 484
+ ],
+ "score": 1.0,
+ "content": "fore, we are free to use deep networks. However, we should pay extra attention to extreme cases",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 495
+ ],
+ "score": 1.0,
+ "content": "where scale changes are of very big magnitude. These are quite rare but still appear in practice.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 493,
+ 505,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 493,
+ 505,
+ 505
+ ],
+ "score": 1.0,
+ "content": "Finally, we see that using SESN with interscale interaction introduces extra equivariance error due",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 505,
+ 505,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 505,
+ 185,
+ 516
+ ],
+ "score": 1.0,
+ "content": "to the truncation of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 185,
+ 505,
+ 192,
+ 514
+ ],
+ "score": 0.76,
+ "content": "S",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 193,
+ 505,
+ 505,
+ 516
+ ],
+ "score": 1.0,
+ "content": ". We will build the networks with either no scale interaction or interaction of 2",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 516,
+ 136,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 516,
+ 136,
+ 527
+ ],
+ "score": 1.0,
+ "content": "scales.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 19.5,
+ "bbox_fs": [
+ 105,
+ 395,
+ 505,
+ 527
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 543,
+ 196,
+ 555
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 542,
+ 198,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 542,
+ 198,
+ 558
+ ],
+ "score": 1.0,
+ "content": "6.2 MNIST-SCALE",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 565,
+ 505,
+ 632
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 566,
+ 505,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 566,
+ 505,
+ 578
+ ],
+ "score": 1.0,
+ "content": "Following Kanazawa et al. (2014); Marcos et al. (2018); Ghosh & Gupta (2019) we conduct experi-",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 505,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 505,
+ 589
+ ],
+ "score": 1.0,
+ "content": "ments on the MNIST-scale dataset. We rescale the images of the MNIST dataset LeCun et al. (1998)",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 588,
+ 505,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 588,
+ 117,
+ 600
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 117,
+ 588,
+ 157,
+ 599
+ ],
+ "score": 0.78,
+ "content": "0 . 3 - 1 . 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 158,
+ 588,
+ 505,
+ 600
+ ],
+ "score": 1.0,
+ "content": "of the original size and pad them with zeros to retain the initial resolution. The scal-",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 599,
+ 505,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 599,
+ 505,
+ 612
+ ],
+ "score": 1.0,
+ "content": "ing factors are sampled uniformly and independently for each image. The obtained dataset is then",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 622
+ ],
+ "score": 1.0,
+ "content": "split into 10,000 for training, 2,000 for evaluation and 50,000 for testing. We generate 6 different",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 621,
+ 287,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 287,
+ 633
+ ],
+ "score": 1.0,
+ "content": "realizations and fix them for all experiments.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 29.5,
+ "bbox_fs": [
+ 105,
+ 566,
+ 505,
+ 633
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 637,
+ 505,
+ 693
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 638,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 638,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "As a baseline model we use the model described in Ghosh & Gupta (2019), which currently holds",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 648,
+ 505,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 505,
+ 662
+ ],
+ "score": 1.0,
+ "content": "the state-of-the-art result on this dataset. It consists of 3 convolutional and 2 fully-connected layers.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 659,
+ 505,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 659,
+ 224,
+ 673
+ ],
+ "score": 1.0,
+ "content": "Each layer has filters of size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 225,
+ 660,
+ 249,
+ 671
+ ],
+ "score": 0.88,
+ "content": "7 \\times 7",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 249,
+ 659,
+ 505,
+ 673
+ ],
+ "score": 1.0,
+ "content": ". We keep the number of trainable parameters almost the same",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 671,
+ 505,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 671,
+ 505,
+ 683
+ ],
+ "score": 1.0,
+ "content": "for all tested methods. This is achieved by varying the number of channels. For scale equivariant",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 681,
+ 384,
+ 694
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 681,
+ 384,
+ 694
+ ],
+ "score": 1.0,
+ "content": "models we add scale projection at the end of the convolutional block.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 105,
+ 638,
+ 505,
+ 694
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 699,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "For SiCNN, DSS, SEVF and our model, we additionally train counterparts where after each con-",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 709,
+ 506,
+ 721
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 506,
+ 721
+ ],
+ "score": 1.0,
+ "content": "volution, an extra projection layer is inserted. Projection layers transform vector features in each",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 733
+ ],
+ "score": 1.0,
+ "content": "spatial position of each channel into scalar ones. All of the layers have now scalar inputs instead of",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": "vector inputs. Therefore, we denote these models with “Scalar”. The original models are denoted as",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 92,
+ 506,
+ 108
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 92,
+ 506,
+ 108
+ ],
+ "score": 1.0,
+ "content": "“Vector”. The exact type of projection depends on the way the vector features are constructed. For",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 117
+ ],
+ "score": 1.0,
+ "content": "SiCNN, DSS, and SESN, we use maximum pooling along the scale dimension, while for SEVF, it",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 287,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 194,
+ 127
+ ],
+ "score": 1.0,
+ "content": "is a calculation of the",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 194,
+ 116,
+ 206,
+ 127
+ ],
+ "score": 0.88,
+ "content": "L _ { 2 }",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 207,
+ 115,
+ 287,
+ 127
+ ],
+ "score": 1.0,
+ "content": "-norm of the vector.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 39,
+ "bbox_fs": [
+ 105,
+ 698,
+ 506,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 126
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 506,
+ 95
+ ],
+ "score": 1.0,
+ "content": "vector inputs. Therefore, we denote these models with “Scalar”. The original models are denoted as",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 92,
+ 506,
+ 108
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 92,
+ 506,
+ 108
+ ],
+ "score": 1.0,
+ "content": "“Vector”. The exact type of projection depends on the way the vector features are constructed. For",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 117
+ ],
+ "score": 1.0,
+ "content": "SiCNN, DSS, and SESN, we use maximum pooling along the scale dimension, while for SEVF, it",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 287,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 194,
+ 127
+ ],
+ "score": 1.0,
+ "content": "is a calculation of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 194,
+ 116,
+ 206,
+ 127
+ ],
+ "score": 0.88,
+ "content": "L _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 207,
+ 115,
+ 287,
+ 127
+ ],
+ "score": 1.0,
+ "content": "-norm of the vector.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 1.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 132,
+ 505,
+ 231
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 131,
+ 506,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 131,
+ 506,
+ 145
+ ],
+ "score": 1.0,
+ "content": "All models are trained with the Adam optimizer Kingma & Ba (2014) for 60 epochs with a batch",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 143,
+ 505,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 143,
+ 505,
+ 155
+ ],
+ "score": 1.0,
+ "content": "size of 128. Initial learning rate is set to 0.01 and divided by 10 after 20 and 40 epochs. We conduct",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 154,
+ 506,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 154,
+ 506,
+ 167
+ ],
+ "score": 1.0,
+ "content": "the experiments with 4 different settings. Following the idea discussed in Ghosh & Gupta (2019),",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 504,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 468,
+ 178
+ ],
+ "score": 1.0,
+ "content": "in addition to the standard setting we train the networks with input images upscaled to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 469,
+ 165,
+ 504,
+ 176
+ ],
+ "score": 0.89,
+ "content": "5 6 \\times 5 6",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 177,
+ 505,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 177,
+ 505,
+ 189
+ ],
+ "score": 1.0,
+ "content": "using bilinear interpolation. This results in all image transformations performed by the network",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 187,
+ 506,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 187,
+ 506,
+ 200
+ ],
+ "score": 1.0,
+ "content": "becoming more stable, which produces less interpolation artifacts. For both input sizes we conduct",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 198,
+ 506,
+ 211
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 198,
+ 506,
+ 211
+ ],
+ "score": 1.0,
+ "content": "the experiments without data augmentation and with scaling augmentation, which results in 4 setups",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 209,
+ 505,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 209,
+ 479,
+ 221
+ ],
+ "score": 1.0,
+ "content": "in total. We run the experiments on 6 different realizations of MNIST-scale and report mean",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 480,
+ 210,
+ 490,
+ 220
+ ],
+ "score": 0.72,
+ "content": "\\pm",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 491,
+ 209,
+ 505,
+ 221
+ ],
+ "score": 1.0,
+ "content": "std",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 220,
+ 214,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 214,
+ 232
+ ],
+ "score": 1.0,
+ "content": "calculated over these runs.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 236,
+ 505,
+ 303
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 236,
+ 505,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 236,
+ 505,
+ 250
+ ],
+ "score": 1.0,
+ "content": "The obtained results are summarized in Table 2. The reported errors may differ a bit from the ones",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 246,
+ 506,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 246,
+ 506,
+ 262
+ ],
+ "score": 1.0,
+ "content": "in the original paper because of the variations in generated datasets and slightly different training",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 258,
+ 506,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 506,
+ 272
+ ],
+ "score": 1.0,
+ "content": "procedure. Nevertheless, we try to keep our configuration as close as possible to Ghosh & Gupta",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 269,
+ 505,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 269,
+ 505,
+ 282
+ ],
+ "score": 1.0,
+ "content": "(2019) which currently demonstrated the best classification accuracy on MNIST-scale. For example,",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 281,
+ 506,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 281,
+ 211,
+ 294
+ ],
+ "score": 1.0,
+ "content": "SS-CNN reports error of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 212,
+ 281,
+ 262,
+ 292
+ ],
+ "score": 0.88,
+ "content": "1 . 9 1 \\pm 0 . 0 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 262,
+ 281,
+ 425,
+ 294
+ ],
+ "score": 1.0,
+ "content": "in Ghosh & Gupta (2019) while it has",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 425,
+ 281,
+ 476,
+ 292
+ ],
+ "score": 0.88,
+ "content": "1 . 8 4 \\pm 0 . 1 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 477,
+ 281,
+ 506,
+ 294
+ ],
+ "score": 1.0,
+ "content": "in our",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 292,
+ 160,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 292,
+ 160,
+ 306
+ ],
+ "score": 1.0,
+ "content": "experiments.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 308,
+ 504,
+ 353
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 308,
+ 505,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 308,
+ 505,
+ 321
+ ],
+ "score": 1.0,
+ "content": "SESN significantly outperforms other methods in all 4 regimes. “Scalar” versions of it already",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 320,
+ 506,
+ 332
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 320,
+ 506,
+ 332
+ ],
+ "score": 1.0,
+ "content": "outperform all previous methods, and “Vector” versions make the gain even more significant. The",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 331,
+ 506,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 331,
+ 506,
+ 343
+ ],
+ "score": 1.0,
+ "content": "global architectures of all models are the same for all rows, which indicates that the way scale",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 342,
+ 284,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 342,
+ 284,
+ 353
+ ],
+ "score": 1.0,
+ "content": "convolution is done plays an important role.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 20.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 368,
+ 165,
+ 379
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 168,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 168,
+ 381
+ ],
+ "score": 1.0,
+ "content": "6.3 STL-10",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 389,
+ 338,
+ 489
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 390,
+ 339,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 390,
+ 339,
+ 401
+ ],
+ "score": 1.0,
+ "content": "In order to evaluate the role of scale equivariance in nat-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 401,
+ 339,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 401,
+ 339,
+ 413
+ ],
+ "score": 1.0,
+ "content": "ural image classification, we conduct the experiments on",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 411,
+ 340,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 411,
+ 340,
+ 423
+ ],
+ "score": 1.0,
+ "content": "STL-10 dataset Coates et al. (2011). This dataset consists",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 423,
+ 339,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 423,
+ 339,
+ 434
+ ],
+ "score": 1.0,
+ "content": "of 8,000 training and 5,000 testing labeled images. Ad-",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 434,
+ 339,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 434,
+ 339,
+ 445
+ ],
+ "score": 1.0,
+ "content": "ditionally, it includes 100,000 unlabeled images. The im-",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 444,
+ 339,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 444,
+ 205,
+ 456
+ ],
+ "score": 1.0,
+ "content": "ages have a resolution of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 445,
+ 235,
+ 455
+ ],
+ "score": 0.91,
+ "content": "9 6 \\times 9 6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 235,
+ 444,
+ 339,
+ 456
+ ],
+ "score": 1.0,
+ "content": "pixels and RGB channels.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 455,
+ 340,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 455,
+ 340,
+ 467
+ ],
+ "score": 1.0,
+ "content": "Labeled images belong to 10 classes such as bird, horse or",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 467,
+ 339,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 467,
+ 339,
+ 478
+ ],
+ "score": 1.0,
+ "content": "car. We use only the labeled subset to demonstrate the",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 478,
+ 310,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 478,
+ 310,
+ 490
+ ],
+ "score": 1.0,
+ "content": "performance of the models in the low data regime.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 28
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 494,
+ 338,
+ 572
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 493,
+ 339,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 493,
+ 339,
+ 507
+ ],
+ "score": 1.0,
+ "content": "The dataset is normalized by subtracting the per-channel",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 505,
+ 339,
+ 517
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 339,
+ 517
+ ],
+ "score": 1.0,
+ "content": "mean and dividing by the per-channel standard deviation.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 516,
+ 339,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 516,
+ 339,
+ 529
+ ],
+ "score": 1.0,
+ "content": "During training, we augment the dataset by applying 12",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 528,
+ 339,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 528,
+ 339,
+ 540
+ ],
+ "score": 1.0,
+ "content": "pixel zero padding and randomly cropping the images to",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 339,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 125,
+ 550
+ ],
+ "score": 1.0,
+ "content": "size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 125,
+ 538,
+ 159,
+ 549
+ ],
+ "score": 0.9,
+ "content": "9 6 \\times 9 6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 160,
+ 538,
+ 339,
+ 550
+ ],
+ "score": 1.0,
+ "content": ". Additionally, random horizontal flips with",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 548,
+ 340,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 548,
+ 152,
+ 562
+ ],
+ "score": 1.0,
+ "content": "probability",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 152,
+ 549,
+ 172,
+ 560
+ ],
+ "score": 0.89,
+ "content": "5 0 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 172,
+ 548,
+ 340,
+ 562
+ ],
+ "score": 1.0,
+ "content": "and Cutout DeVries & Taylor (2017) with",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 559,
+ 221,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 559,
+ 221,
+ 573
+ ],
+ "score": 1.0,
+ "content": "1 hole of 32 pixels are used.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 577,
+ 340,
+ 643
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 576,
+ 339,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 576,
+ 339,
+ 589
+ ],
+ "score": 1.0,
+ "content": "As a baseline we choose WideResNet Zagoruyko & Ko-",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 588,
+ 339,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 588,
+ 339,
+ 600
+ ],
+ "score": 1.0,
+ "content": "modakis (2016) with 16 layers and a widening factor of 8.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 599,
+ 339,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 339,
+ 611
+ ],
+ "score": 1.0,
+ "content": "We set dropout probability to 0.3 in all blocks. We train",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 610,
+ 340,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 610,
+ 340,
+ 622
+ ],
+ "score": 1.0,
+ "content": "SESN-A with just vector features. For SESN-B we use",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 621,
+ 339,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 339,
+ 632
+ ],
+ "score": 1.0,
+ "content": "maximum scalar projection several times in the interme-",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 631,
+ 342,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 631,
+ 342,
+ 644
+ ],
+ "score": 1.0,
+ "content": "diate layers, and for SESN-C we use interscale interaction.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 42.5
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 347,
+ 408,
+ 504,
+ 541
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 347,
+ 408,
+ 504,
+ 541
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 347,
+ 408,
+ 504,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 347,
+ 408,
+ 504,
+ 541
+ ],
+ "score": 0.981,
+ "html": "| Method | Error, % | #Params |
| WRN | 11.48 | 11.0 M |
| SiCNN | 11.62 | 11.0 M |
| SI-ConvNet | 12.48 | 11.0 M |
| DSS | 11.28 | 11.0 M |
| SS-CNN | 25.47 | 10.8 M |
| SESN-A | 10.83 | 11.0 M |
| SESN-B | 8.51 | 11.0 M |
| SESN-C | 14.08 | 11.0 M |
| Harm WRN | 9.55 | 11.0 M |
",
+ "type": "table",
+ "image_path": "4d32f05f49eec5a1aa266914ff213cdd1b415b4908692d90c08c7b8e0fe5a749.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 50.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 347,
+ 408,
+ 504,
+ 421.3
+ ],
+ "spans": [],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 347,
+ 421.3,
+ 504,
+ 434.6
+ ],
+ "spans": [],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 347,
+ 434.6,
+ 504,
+ 447.90000000000003
+ ],
+ "spans": [],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 347,
+ 447.90000000000003,
+ 504,
+ 461.20000000000005
+ ],
+ "spans": [],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 347,
+ 461.20000000000005,
+ 504,
+ 474.50000000000006
+ ],
+ "spans": [],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 347,
+ 474.50000000000006,
+ 504,
+ 487.80000000000007
+ ],
+ "spans": [],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 347,
+ 487.80000000000007,
+ 504,
+ 501.1000000000001
+ ],
+ "spans": [],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 347,
+ 501.1000000000001,
+ 504,
+ 514.4000000000001
+ ],
+ "spans": [],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 347,
+ 514.4000000000001,
+ 504,
+ 527.7
+ ],
+ "spans": [],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 347,
+ 527.7,
+ 504,
+ 541.0
+ ],
+ "spans": [],
+ "index": 55
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 346,
+ 553,
+ 504,
+ 609
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 346,
+ 553,
+ 473,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 346,
+ 553,
+ 473,
+ 565
+ ],
+ "score": 1.0,
+ "content": "Table 3: Classification error on",
+ "type": "text"
+ }
+ ],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 345,
+ 564,
+ 505,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 345,
+ 564,
+ 505,
+ 577
+ ],
+ "score": 1.0,
+ "content": "STL-10. The best results are bold.",
+ "type": "text"
+ }
+ ],
+ "index": 57
+ },
+ {
+ "bbox": [
+ 346,
+ 575,
+ 505,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 346,
+ 575,
+ 505,
+ 587
+ ],
+ "score": 1.0,
+ "content": "We additionally report the current best",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 345,
+ 586,
+ 505,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 345,
+ 586,
+ 505,
+ 598
+ ],
+ "score": 1.0,
+ "content": "result achieved by Harm WRN from",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ },
+ {
+ "bbox": [
+ 346,
+ 597,
+ 429,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 346,
+ 597,
+ 429,
+ 610
+ ],
+ "score": 1.0,
+ "content": "Ulicny et al. (2019).",
+ "type": "text"
+ }
+ ],
+ "index": 60
+ }
+ ],
+ "index": 58
+ }
+ ],
+ "index": 54.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 649,
+ 500,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 502,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 502,
+ 662
+ ],
+ "score": 1.0,
+ "content": "All models are trained for 1000 epochs with a batch size of 128. We use SGD optimizer with",
+ "type": "text"
+ }
+ ],
+ "index": 61
+ },
+ {
+ "bbox": [
+ 105,
+ 659,
+ 502,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 659,
+ 303,
+ 673
+ ],
+ "score": 1.0,
+ "content": "Nesterov momentum of 0.9 and weight decay of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 303,
+ 659,
+ 337,
+ 670
+ ],
+ "score": 0.9,
+ "content": "5 \\cdot 1 0 ^ { - 4 } ",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 337,
+ 659,
+ 502,
+ 673
+ ],
+ "score": 1.0,
+ "content": ". The initial learning rate is set to 0.1 and",
+ "type": "text"
+ }
+ ],
+ "index": 62
+ },
+ {
+ "bbox": [
+ 106,
+ 671,
+ 303,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 671,
+ 303,
+ 683
+ ],
+ "score": 1.0,
+ "content": "divided by 5 after 300, 400, 600 and 800 epochs.",
+ "type": "text"
+ }
+ ],
+ "index": 63
+ }
+ ],
+ "index": 62
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "The results are summarized in Table 3. We found SEVF training unstable and therefore do not",
+ "type": "text"
+ }
+ ],
+ "index": 64
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "include it in the table. Pure scale-invariant SI-ConvNet and SS-CNN demonstrate significantly worse",
+ "type": "text"
+ }
+ ],
+ "index": 65
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "results than the baseline. We note the importance of equivariance for deep networks. We also find",
+ "type": "text"
+ }
+ ],
+ "index": 66
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "score": 1.0,
+ "content": "that SESN-C performs significantly worse than SESN-A and SESN-B due to high equivariance",
+ "type": "text"
+ }
+ ],
+ "index": 67
+ }
+ ],
+ "index": 65.5
+ }
+ ],
+ "page_idx": 7,
+ "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 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 761
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 761
+ ],
+ "score": 1.0,
+ "content": "8",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 126
+ ],
+ "lines": [],
+ "index": 1.5,
+ "bbox_fs": [
+ 105,
+ 82,
+ 506,
+ 127
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 132,
+ 505,
+ 231
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 131,
+ 506,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 131,
+ 506,
+ 145
+ ],
+ "score": 1.0,
+ "content": "All models are trained with the Adam optimizer Kingma & Ba (2014) for 60 epochs with a batch",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 143,
+ 505,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 143,
+ 505,
+ 155
+ ],
+ "score": 1.0,
+ "content": "size of 128. Initial learning rate is set to 0.01 and divided by 10 after 20 and 40 epochs. We conduct",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 154,
+ 506,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 154,
+ 506,
+ 167
+ ],
+ "score": 1.0,
+ "content": "the experiments with 4 different settings. Following the idea discussed in Ghosh & Gupta (2019),",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 504,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 468,
+ 178
+ ],
+ "score": 1.0,
+ "content": "in addition to the standard setting we train the networks with input images upscaled to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 469,
+ 165,
+ 504,
+ 176
+ ],
+ "score": 0.89,
+ "content": "5 6 \\times 5 6",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 177,
+ 505,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 177,
+ 505,
+ 189
+ ],
+ "score": 1.0,
+ "content": "using bilinear interpolation. This results in all image transformations performed by the network",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 187,
+ 506,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 187,
+ 506,
+ 200
+ ],
+ "score": 1.0,
+ "content": "becoming more stable, which produces less interpolation artifacts. For both input sizes we conduct",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 198,
+ 506,
+ 211
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 198,
+ 506,
+ 211
+ ],
+ "score": 1.0,
+ "content": "the experiments without data augmentation and with scaling augmentation, which results in 4 setups",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 209,
+ 505,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 209,
+ 479,
+ 221
+ ],
+ "score": 1.0,
+ "content": "in total. We run the experiments on 6 different realizations of MNIST-scale and report mean",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 480,
+ 210,
+ 490,
+ 220
+ ],
+ "score": 0.72,
+ "content": "\\pm",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 491,
+ 209,
+ 505,
+ 221
+ ],
+ "score": 1.0,
+ "content": "std",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 220,
+ 214,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 220,
+ 214,
+ 232
+ ],
+ "score": 1.0,
+ "content": "calculated over these runs.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 8,
+ "bbox_fs": [
+ 105,
+ 131,
+ 506,
+ 232
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 236,
+ 505,
+ 303
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 236,
+ 505,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 236,
+ 505,
+ 250
+ ],
+ "score": 1.0,
+ "content": "The obtained results are summarized in Table 2. The reported errors may differ a bit from the ones",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 246,
+ 506,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 246,
+ 506,
+ 262
+ ],
+ "score": 1.0,
+ "content": "in the original paper because of the variations in generated datasets and slightly different training",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 258,
+ 506,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 258,
+ 506,
+ 272
+ ],
+ "score": 1.0,
+ "content": "procedure. Nevertheless, we try to keep our configuration as close as possible to Ghosh & Gupta",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 269,
+ 505,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 269,
+ 505,
+ 282
+ ],
+ "score": 1.0,
+ "content": "(2019) which currently demonstrated the best classification accuracy on MNIST-scale. For example,",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 281,
+ 506,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 281,
+ 211,
+ 294
+ ],
+ "score": 1.0,
+ "content": "SS-CNN reports error of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 212,
+ 281,
+ 262,
+ 292
+ ],
+ "score": 0.88,
+ "content": "1 . 9 1 \\pm 0 . 0 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 262,
+ 281,
+ 425,
+ 294
+ ],
+ "score": 1.0,
+ "content": "in Ghosh & Gupta (2019) while it has",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 425,
+ 281,
+ 476,
+ 292
+ ],
+ "score": 0.88,
+ "content": "1 . 8 4 \\pm 0 . 1 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 477,
+ 281,
+ 506,
+ 294
+ ],
+ "score": 1.0,
+ "content": "in our",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 292,
+ 160,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 292,
+ 160,
+ 306
+ ],
+ "score": 1.0,
+ "content": "experiments.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 15.5,
+ "bbox_fs": [
+ 105,
+ 236,
+ 506,
+ 306
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 308,
+ 504,
+ 353
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 308,
+ 505,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 308,
+ 505,
+ 321
+ ],
+ "score": 1.0,
+ "content": "SESN significantly outperforms other methods in all 4 regimes. “Scalar” versions of it already",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 320,
+ 506,
+ 332
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 320,
+ 506,
+ 332
+ ],
+ "score": 1.0,
+ "content": "outperform all previous methods, and “Vector” versions make the gain even more significant. The",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 331,
+ 506,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 331,
+ 506,
+ 343
+ ],
+ "score": 1.0,
+ "content": "global architectures of all models are the same for all rows, which indicates that the way scale",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 342,
+ 284,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 342,
+ 284,
+ 353
+ ],
+ "score": 1.0,
+ "content": "convolution is done plays an important role.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 20.5,
+ "bbox_fs": [
+ 105,
+ 308,
+ 506,
+ 353
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 368,
+ 165,
+ 379
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 168,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 168,
+ 381
+ ],
+ "score": 1.0,
+ "content": "6.3 STL-10",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 389,
+ 338,
+ 489
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 390,
+ 339,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 390,
+ 339,
+ 401
+ ],
+ "score": 1.0,
+ "content": "In order to evaluate the role of scale equivariance in nat-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 401,
+ 339,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 401,
+ 339,
+ 413
+ ],
+ "score": 1.0,
+ "content": "ural image classification, we conduct the experiments on",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 411,
+ 340,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 411,
+ 340,
+ 423
+ ],
+ "score": 1.0,
+ "content": "STL-10 dataset Coates et al. (2011). This dataset consists",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 423,
+ 339,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 423,
+ 339,
+ 434
+ ],
+ "score": 1.0,
+ "content": "of 8,000 training and 5,000 testing labeled images. Ad-",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 434,
+ 339,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 434,
+ 339,
+ 445
+ ],
+ "score": 1.0,
+ "content": "ditionally, it includes 100,000 unlabeled images. The im-",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 444,
+ 339,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 444,
+ 205,
+ 456
+ ],
+ "score": 1.0,
+ "content": "ages have a resolution of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 205,
+ 445,
+ 235,
+ 455
+ ],
+ "score": 0.91,
+ "content": "9 6 \\times 9 6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 235,
+ 444,
+ 339,
+ 456
+ ],
+ "score": 1.0,
+ "content": "pixels and RGB channels.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 455,
+ 340,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 455,
+ 340,
+ 467
+ ],
+ "score": 1.0,
+ "content": "Labeled images belong to 10 classes such as bird, horse or",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 467,
+ 339,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 467,
+ 339,
+ 478
+ ],
+ "score": 1.0,
+ "content": "car. We use only the labeled subset to demonstrate the",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 478,
+ 310,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 478,
+ 310,
+ 490
+ ],
+ "score": 1.0,
+ "content": "performance of the models in the low data regime.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 28,
+ "bbox_fs": [
+ 105,
+ 390,
+ 340,
+ 490
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 494,
+ 338,
+ 572
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 493,
+ 339,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 493,
+ 339,
+ 507
+ ],
+ "score": 1.0,
+ "content": "The dataset is normalized by subtracting the per-channel",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 505,
+ 339,
+ 517
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 339,
+ 517
+ ],
+ "score": 1.0,
+ "content": "mean and dividing by the per-channel standard deviation.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 516,
+ 339,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 516,
+ 339,
+ 529
+ ],
+ "score": 1.0,
+ "content": "During training, we augment the dataset by applying 12",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 528,
+ 339,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 528,
+ 339,
+ 540
+ ],
+ "score": 1.0,
+ "content": "pixel zero padding and randomly cropping the images to",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 339,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 125,
+ 550
+ ],
+ "score": 1.0,
+ "content": "size",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 125,
+ 538,
+ 159,
+ 549
+ ],
+ "score": 0.9,
+ "content": "9 6 \\times 9 6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 160,
+ 538,
+ 339,
+ 550
+ ],
+ "score": 1.0,
+ "content": ". Additionally, random horizontal flips with",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 548,
+ 340,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 548,
+ 152,
+ 562
+ ],
+ "score": 1.0,
+ "content": "probability",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 152,
+ 549,
+ 172,
+ 560
+ ],
+ "score": 0.89,
+ "content": "5 0 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 172,
+ 548,
+ 340,
+ 562
+ ],
+ "score": 1.0,
+ "content": "and Cutout DeVries & Taylor (2017) with",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 559,
+ 221,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 559,
+ 221,
+ 573
+ ],
+ "score": 1.0,
+ "content": "1 hole of 32 pixels are used.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 36,
+ "bbox_fs": [
+ 105,
+ 493,
+ 340,
+ 573
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 577,
+ 340,
+ 643
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 576,
+ 339,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 576,
+ 339,
+ 589
+ ],
+ "score": 1.0,
+ "content": "As a baseline we choose WideResNet Zagoruyko & Ko-",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 588,
+ 339,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 588,
+ 339,
+ 600
+ ],
+ "score": 1.0,
+ "content": "modakis (2016) with 16 layers and a widening factor of 8.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 599,
+ 339,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 339,
+ 611
+ ],
+ "score": 1.0,
+ "content": "We set dropout probability to 0.3 in all blocks. We train",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 610,
+ 340,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 610,
+ 340,
+ 622
+ ],
+ "score": 1.0,
+ "content": "SESN-A with just vector features. For SESN-B we use",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 621,
+ 339,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 339,
+ 632
+ ],
+ "score": 1.0,
+ "content": "maximum scalar projection several times in the interme-",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 631,
+ 342,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 631,
+ 342,
+ 644
+ ],
+ "score": 1.0,
+ "content": "diate layers, and for SESN-C we use interscale interaction.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 42.5,
+ "bbox_fs": [
+ 105,
+ 576,
+ 342,
+ 644
+ ]
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 347,
+ 408,
+ 504,
+ 541
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 347,
+ 408,
+ 504,
+ 541
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 347,
+ 408,
+ 504,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 347,
+ 408,
+ 504,
+ 541
+ ],
+ "score": 0.981,
+ "html": "| Method | Error, % | #Params |
| WRN | 11.48 | 11.0 M |
| SiCNN | 11.62 | 11.0 M |
| SI-ConvNet | 12.48 | 11.0 M |
| DSS | 11.28 | 11.0 M |
| SS-CNN | 25.47 | 10.8 M |
| SESN-A | 10.83 | 11.0 M |
| SESN-B | 8.51 | 11.0 M |
| SESN-C | 14.08 | 11.0 M |
| Harm WRN | 9.55 | 11.0 M |
",
+ "type": "table",
+ "image_path": "4d32f05f49eec5a1aa266914ff213cdd1b415b4908692d90c08c7b8e0fe5a749.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 50.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 347,
+ 408,
+ 504,
+ 421.3
+ ],
+ "spans": [],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 347,
+ 421.3,
+ 504,
+ 434.6
+ ],
+ "spans": [],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 347,
+ 434.6,
+ 504,
+ 447.90000000000003
+ ],
+ "spans": [],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 347,
+ 447.90000000000003,
+ 504,
+ 461.20000000000005
+ ],
+ "spans": [],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 347,
+ 461.20000000000005,
+ 504,
+ 474.50000000000006
+ ],
+ "spans": [],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 347,
+ 474.50000000000006,
+ 504,
+ 487.80000000000007
+ ],
+ "spans": [],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 347,
+ 487.80000000000007,
+ 504,
+ 501.1000000000001
+ ],
+ "spans": [],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 347,
+ 501.1000000000001,
+ 504,
+ 514.4000000000001
+ ],
+ "spans": [],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 347,
+ 514.4000000000001,
+ 504,
+ 527.7
+ ],
+ "spans": [],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 347,
+ 527.7,
+ 504,
+ 541.0
+ ],
+ "spans": [],
+ "index": 55
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 346,
+ 553,
+ 504,
+ 609
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 346,
+ 553,
+ 473,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 346,
+ 553,
+ 473,
+ 565
+ ],
+ "score": 1.0,
+ "content": "Table 3: Classification error on",
+ "type": "text"
+ }
+ ],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 345,
+ 564,
+ 505,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 345,
+ 564,
+ 505,
+ 577
+ ],
+ "score": 1.0,
+ "content": "STL-10. The best results are bold.",
+ "type": "text"
+ }
+ ],
+ "index": 57
+ },
+ {
+ "bbox": [
+ 346,
+ 575,
+ 505,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 346,
+ 575,
+ 505,
+ 587
+ ],
+ "score": 1.0,
+ "content": "We additionally report the current best",
+ "type": "text"
+ }
+ ],
+ "index": 58
+ },
+ {
+ "bbox": [
+ 345,
+ 586,
+ 505,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 345,
+ 586,
+ 505,
+ 598
+ ],
+ "score": 1.0,
+ "content": "result achieved by Harm WRN from",
+ "type": "text"
+ }
+ ],
+ "index": 59
+ },
+ {
+ "bbox": [
+ 346,
+ 597,
+ 429,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 346,
+ 597,
+ 429,
+ 610
+ ],
+ "score": 1.0,
+ "content": "Ulicny et al. (2019).",
+ "type": "text"
+ }
+ ],
+ "index": 60
+ }
+ ],
+ "index": 58
+ }
+ ],
+ "index": 54.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 649,
+ 500,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 502,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 502,
+ 662
+ ],
+ "score": 1.0,
+ "content": "All models are trained for 1000 epochs with a batch size of 128. We use SGD optimizer with",
+ "type": "text"
+ }
+ ],
+ "index": 61
+ },
+ {
+ "bbox": [
+ 105,
+ 659,
+ 502,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 659,
+ 303,
+ 673
+ ],
+ "score": 1.0,
+ "content": "Nesterov momentum of 0.9 and weight decay of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 303,
+ 659,
+ 337,
+ 670
+ ],
+ "score": 0.9,
+ "content": "5 \\cdot 1 0 ^ { - 4 } ",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 337,
+ 659,
+ 502,
+ 673
+ ],
+ "score": 1.0,
+ "content": ". The initial learning rate is set to 0.1 and",
+ "type": "text"
+ }
+ ],
+ "index": 62
+ },
+ {
+ "bbox": [
+ 106,
+ 671,
+ 303,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 671,
+ 303,
+ 683
+ ],
+ "score": 1.0,
+ "content": "divided by 5 after 300, 400, 600 and 800 epochs.",
+ "type": "text"
+ }
+ ],
+ "index": 63
+ }
+ ],
+ "index": 62,
+ "bbox_fs": [
+ 105,
+ 647,
+ 502,
+ 683
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "The results are summarized in Table 3. We found SEVF training unstable and therefore do not",
+ "type": "text"
+ }
+ ],
+ "index": 64
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "include it in the table. Pure scale-invariant SI-ConvNet and SS-CNN demonstrate significantly worse",
+ "type": "text"
+ }
+ ],
+ "index": 65
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "results than the baseline. We note the importance of equivariance for deep networks. We also find",
+ "type": "text"
+ }
+ ],
+ "index": 66
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "score": 1.0,
+ "content": "that SESN-C performs significantly worse than SESN-A and SESN-B due to high equivariance",
+ "type": "text"
+ }
+ ],
+ "index": 67
+ },
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "error caused by interscale interaction. SESN-B significantly improves the results of both WRN",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "and DSS due to the projection between scales. The maximum scale projection makes the weights",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 117
+ ],
+ "score": 1.0,
+ "content": "of the next layer to have a maximum receptive field in the space of scales. This is an easy yet",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 127
+ ],
+ "score": 1.0,
+ "content": "effective method for capturing the correlations between different scales. This experiment shows that",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 127,
+ 505,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 127,
+ 505,
+ 138
+ ],
+ "score": 1.0,
+ "content": "scale-equivariance is a very useful inductive bias for natural image classification with deep neural",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 148,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 148,
+ 149
+ ],
+ "score": 1.0,
+ "content": "networks.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 65.5,
+ "bbox_fs": [
+ 105,
+ 687,
+ 506,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 82,
+ 505,
+ 148
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "error caused by interscale interaction. SESN-B significantly improves the results of both WRN",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 93,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "and DSS due to the projection between scales. The maximum scale projection makes the weights",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 104,
+ 506,
+ 117
+ ],
+ "score": 1.0,
+ "content": "of the next layer to have a maximum receptive field in the space of scales. This is an easy yet",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 115,
+ 506,
+ 127
+ ],
+ "score": 1.0,
+ "content": "effective method for capturing the correlations between different scales. This experiment shows that",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 127,
+ 505,
+ 138
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 127,
+ 505,
+ 138
+ ],
+ "score": 1.0,
+ "content": "scale-equivariance is a very useful inductive bias for natural image classification with deep neural",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 137,
+ 148,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 137,
+ 148,
+ 149
+ ],
+ "score": 1.0,
+ "content": "networks.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 154,
+ 505,
+ 198
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 153,
+ 505,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 153,
+ 505,
+ 167
+ ],
+ "score": 1.0,
+ "content": "To the best of our knowledge, the proposed method achieves a new state-of-the-art result on the",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 177
+ ],
+ "score": 1.0,
+ "content": "STL-10 dataset in the supervised learning setting. The previous lowest error is demonstrated in",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 176,
+ 505,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 176,
+ 505,
+ 188
+ ],
+ "score": 1.0,
+ "content": "Ulicny et al. (2019). The authors propose Harm WRN — a network where the convolutional kernels",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 187,
+ 413,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 413,
+ 199
+ ],
+ "score": 1.0,
+ "content": "are represented as a linear combination of Discrete Cosine Transform filters.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 214,
+ 190,
+ 227
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 212,
+ 192,
+ 230
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 212,
+ 192,
+ 230
+ ],
+ "score": 1.0,
+ "content": "7 DISCUSSION",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 239,
+ 505,
+ 306
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 239,
+ 505,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 239,
+ 505,
+ 252
+ ],
+ "score": 1.0,
+ "content": "In this paper, we have presented the theory of Scale-Equivariant Steerable Networks. We started",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 250,
+ 505,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 250,
+ 505,
+ 262
+ ],
+ "score": 1.0,
+ "content": "from the scaling transformation and its application to continuous functions. We have obtained",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 260,
+ 506,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 260,
+ 506,
+ 275
+ ],
+ "score": 1.0,
+ "content": "the exact formula for scale-equivariant mappings and demonstrated how it can be implemented",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 272,
+ 505,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 272,
+ 505,
+ 284
+ ],
+ "score": 1.0,
+ "content": "for discretized signals. We have demonstrated that this approach outperforms other methods for",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 283,
+ 505,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 283,
+ 505,
+ 295
+ ],
+ "score": 1.0,
+ "content": "scale-equivariant and local scale-invariant CNNs. It demonstrated new state-of-the-art results on",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 293,
+ 406,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 406,
+ 308
+ ],
+ "score": 1.0,
+ "content": "MNIST-scale and on the STL-10 dataset in the supervised learning setting.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 13.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 311,
+ 504,
+ 355
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 311,
+ 505,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 311,
+ 505,
+ 324
+ ],
+ "score": 1.0,
+ "content": "We suppose that the most exciting possible application of SESN is in computer vision for au-",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 322,
+ 506,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 322,
+ 506,
+ 334
+ ],
+ "score": 1.0,
+ "content": "tonomous vehicles. Rapidly changing distances between the objects cause significant scale vari-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 333,
+ 506,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 333,
+ 506,
+ 345
+ ],
+ "score": 1.0,
+ "content": "ations which makes this well suited for our work. We especially highlight the direction of siamese",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 344,
+ 472,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 344,
+ 472,
+ 357
+ ],
+ "score": 1.0,
+ "content": "visual tracking where the equivariance to principle transformations plays an important role.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 18.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 368,
+ 200,
+ 378
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 369,
+ 200,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 369,
+ 200,
+ 379
+ ],
+ "score": 1.0,
+ "content": "ACKNOWLEDGMENTS",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 387,
+ 504,
+ 408
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 386,
+ 505,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 386,
+ 505,
+ 399
+ ],
+ "score": 1.0,
+ "content": "We thank Daniel Worrall for insightful discussion, Thomas Andy Keller, Victor Garcia, Artem",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 397,
+ 371,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 397,
+ 371,
+ 409
+ ],
+ "score": 1.0,
+ "content": "Moskalev and Konrad Groh for valuable comments and feedback.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 425,
+ 175,
+ 437
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 425,
+ 176,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 425,
+ 176,
+ 439
+ ],
+ "score": 1.0,
+ "content": "REFERENCES",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 443,
+ 504,
+ 466
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 443,
+ 506,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 443,
+ 506,
+ 456
+ ],
+ "score": 1.0,
+ "content": "Etienne Barnard and David Casasent. Invariance and neural nets. IEEE Transactions on neural",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 115,
+ 455,
+ 242,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 455,
+ 242,
+ 466
+ ],
+ "score": 1.0,
+ "content": "networks, 2(5):498–508, 1991.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 25.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 473,
+ 504,
+ 507
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 472,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 472,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 116,
+ 484,
+ 504,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 484,
+ 504,
+ 497
+ ],
+ "score": 1.0,
+ "content": "feature learning. In Proceedings of the fourteenth international conference on artificial intelli-",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 496,
+ 279,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 496,
+ 279,
+ 507
+ ],
+ "score": 1.0,
+ "content": "gence and statistics, pp. 215–223, 2011.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 513,
+ 503,
+ 537
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 513,
+ 504,
+ 526
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 513,
+ 504,
+ 526
+ ],
+ "score": 1.0,
+ "content": "Taco Cohen and Max Welling. Group equivariant convolutional networks. In International confer-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 115,
+ 525,
+ 319,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 525,
+ 319,
+ 537
+ ],
+ "score": 1.0,
+ "content": "ence on machine learning, pp. 2990–2999, 2016a.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 542,
+ 504,
+ 566
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 543,
+ 505,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 543,
+ 505,
+ 556
+ ],
+ "score": 1.0,
+ "content": "Taco Cohen, Mario Geiger, Jonas Kohler, and Max Welling. Convolutional networks for spherical ¨",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 554,
+ 312,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 554,
+ 312,
+ 567
+ ],
+ "score": 1.0,
+ "content": "signals. arXiv preprint arXiv:1709.04893, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 572,
+ 470,
+ 585
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 471,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 471,
+ 586
+ ],
+ "score": 1.0,
+ "content": "Taco S Cohen and Max Welling. Steerable cnns. arXiv preprint arXiv:1612.08498, 2016b.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 591,
+ 502,
+ 614
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 591,
+ 504,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 591,
+ 504,
+ 604
+ ],
+ "score": 1.0,
+ "content": "Taco S Cohen, Maurice Weiler, Berkay Kicanaoglu, and Max Welling. Gauge equivariant convolu-",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 116,
+ 602,
+ 443,
+ 615
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 602,
+ 443,
+ 615
+ ],
+ "score": 1.0,
+ "content": "tional networks and the icosahedral cnn. arXiv preprint arXiv:1902.04615, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 621,
+ 503,
+ 644
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 621,
+ 505,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 621,
+ 505,
+ 633
+ ],
+ "score": 1.0,
+ "content": "Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 116,
+ 632,
+ 329,
+ 643
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 632,
+ 329,
+ 643
+ ],
+ "score": 1.0,
+ "content": "with cutout. arXiv preprint arXiv:1708.04552, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 37.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 650,
+ 505,
+ 684
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 505,
+ 664
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 505,
+ 664
+ ],
+ "score": 1.0,
+ "content": "Carlos Esteves, Christine Allen-Blanchette, Ameesh Makadia, and Kostas Daniilidis. Learning so",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 116,
+ 661,
+ 505,
+ 675
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 661,
+ 505,
+ 675
+ ],
+ "score": 1.0,
+ "content": "(3) equivariant representations with spherical cnns. In Proceedings of the European Conference",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 116,
+ 672,
+ 305,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 672,
+ 305,
+ 685
+ ],
+ "score": 1.0,
+ "content": "on Computer Vision (ECCV), pp. 52–68, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 690,
+ 471,
+ 703
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 690,
+ 471,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 690,
+ 471,
+ 703
+ ],
+ "score": 1.0,
+ "content": "Gerald B Folland. A course in abstract harmonic analysis. Chapman and Hall/CRC, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 42
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 709,
+ 502,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 708,
+ 504,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 708,
+ 504,
+ 722
+ ],
+ "score": 1.0,
+ "content": "William T. Freeman and Edward H Adelson. The design and use of steerable filters. IEEE Transac-",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 115,
+ 721,
+ 398,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 721,
+ 398,
+ 732
+ ],
+ "score": 1.0,
+ "content": "tions on Pattern Analysis & Machine Intelligence, (9):891–906, 1991.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43.5
+ }
+ ],
+ "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 2020",
+ "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": [
+ 107,
+ 82,
+ 505,
+ 148
+ ],
+ "lines": [],
+ "index": 2.5,
+ "bbox_fs": [
+ 105,
+ 82,
+ 506,
+ 149
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 154,
+ 505,
+ 198
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 153,
+ 505,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 153,
+ 505,
+ 167
+ ],
+ "score": 1.0,
+ "content": "To the best of our knowledge, the proposed method achieves a new state-of-the-art result on the",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 165,
+ 505,
+ 177
+ ],
+ "score": 1.0,
+ "content": "STL-10 dataset in the supervised learning setting. The previous lowest error is demonstrated in",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 176,
+ 505,
+ 188
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 176,
+ 505,
+ 188
+ ],
+ "score": 1.0,
+ "content": "Ulicny et al. (2019). The authors propose Harm WRN — a network where the convolutional kernels",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 187,
+ 413,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 187,
+ 413,
+ 199
+ ],
+ "score": 1.0,
+ "content": "are represented as a linear combination of Discrete Cosine Transform filters.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 7.5,
+ "bbox_fs": [
+ 105,
+ 153,
+ 505,
+ 199
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 214,
+ 190,
+ 227
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 212,
+ 192,
+ 230
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 212,
+ 192,
+ 230
+ ],
+ "score": 1.0,
+ "content": "7 DISCUSSION",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 239,
+ 505,
+ 306
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 239,
+ 505,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 239,
+ 505,
+ 252
+ ],
+ "score": 1.0,
+ "content": "In this paper, we have presented the theory of Scale-Equivariant Steerable Networks. We started",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 250,
+ 505,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 250,
+ 505,
+ 262
+ ],
+ "score": 1.0,
+ "content": "from the scaling transformation and its application to continuous functions. We have obtained",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 260,
+ 506,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 260,
+ 506,
+ 275
+ ],
+ "score": 1.0,
+ "content": "the exact formula for scale-equivariant mappings and demonstrated how it can be implemented",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 272,
+ 505,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 272,
+ 505,
+ 284
+ ],
+ "score": 1.0,
+ "content": "for discretized signals. We have demonstrated that this approach outperforms other methods for",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 283,
+ 505,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 283,
+ 505,
+ 295
+ ],
+ "score": 1.0,
+ "content": "scale-equivariant and local scale-invariant CNNs. It demonstrated new state-of-the-art results on",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 293,
+ 406,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 406,
+ 308
+ ],
+ "score": 1.0,
+ "content": "MNIST-scale and on the STL-10 dataset in the supervised learning setting.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 13.5,
+ "bbox_fs": [
+ 105,
+ 239,
+ 506,
+ 308
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 311,
+ 504,
+ 355
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 311,
+ 505,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 311,
+ 505,
+ 324
+ ],
+ "score": 1.0,
+ "content": "We suppose that the most exciting possible application of SESN is in computer vision for au-",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 322,
+ 506,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 322,
+ 506,
+ 334
+ ],
+ "score": 1.0,
+ "content": "tonomous vehicles. Rapidly changing distances between the objects cause significant scale vari-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 333,
+ 506,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 333,
+ 506,
+ 345
+ ],
+ "score": 1.0,
+ "content": "ations which makes this well suited for our work. We especially highlight the direction of siamese",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 344,
+ 472,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 344,
+ 472,
+ 357
+ ],
+ "score": 1.0,
+ "content": "visual tracking where the equivariance to principle transformations plays an important role.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 18.5,
+ "bbox_fs": [
+ 105,
+ 311,
+ 506,
+ 357
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 368,
+ 200,
+ 378
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 369,
+ 200,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 369,
+ 200,
+ 379
+ ],
+ "score": 1.0,
+ "content": "ACKNOWLEDGMENTS",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 387,
+ 504,
+ 408
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 386,
+ 505,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 386,
+ 505,
+ 399
+ ],
+ "score": 1.0,
+ "content": "We thank Daniel Worrall for insightful discussion, Thomas Andy Keller, Victor Garcia, Artem",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 397,
+ 371,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 397,
+ 371,
+ 409
+ ],
+ "score": 1.0,
+ "content": "Moskalev and Konrad Groh for valuable comments and feedback.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.5,
+ "bbox_fs": [
+ 106,
+ 386,
+ 505,
+ 409
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 425,
+ 175,
+ 437
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 425,
+ 176,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 425,
+ 176,
+ 439
+ ],
+ "score": 1.0,
+ "content": "REFERENCES",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 443,
+ 504,
+ 466
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 443,
+ 506,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 443,
+ 506,
+ 456
+ ],
+ "score": 1.0,
+ "content": "Etienne Barnard and David Casasent. Invariance and neural nets. IEEE Transactions on neural",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 115,
+ 455,
+ 242,
+ 466
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 455,
+ 242,
+ 466
+ ],
+ "score": 1.0,
+ "content": "networks, 2(5):498–508, 1991.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 25.5,
+ "bbox_fs": [
+ 105,
+ 443,
+ 506,
+ 466
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 473,
+ 504,
+ 507
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 472,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 472,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 116,
+ 484,
+ 504,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 484,
+ 504,
+ 497
+ ],
+ "score": 1.0,
+ "content": "feature learning. In Proceedings of the fourteenth international conference on artificial intelli-",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 496,
+ 279,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 496,
+ 279,
+ 507
+ ],
+ "score": 1.0,
+ "content": "gence and statistics, pp. 215–223, 2011.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28,
+ "bbox_fs": [
+ 105,
+ 472,
+ 505,
+ 507
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 513,
+ 503,
+ 537
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 513,
+ 504,
+ 526
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 513,
+ 504,
+ 526
+ ],
+ "score": 1.0,
+ "content": "Taco Cohen and Max Welling. Group equivariant convolutional networks. In International confer-",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 115,
+ 525,
+ 319,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 525,
+ 319,
+ 537
+ ],
+ "score": 1.0,
+ "content": "ence on machine learning, pp. 2990–2999, 2016a.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.5,
+ "bbox_fs": [
+ 106,
+ 513,
+ 504,
+ 537
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 542,
+ 504,
+ 566
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 543,
+ 505,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 543,
+ 505,
+ 556
+ ],
+ "score": 1.0,
+ "content": "Taco Cohen, Mario Geiger, Jonas Kohler, and Max Welling. Convolutional networks for spherical ¨",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 554,
+ 312,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 554,
+ 312,
+ 567
+ ],
+ "score": 1.0,
+ "content": "signals. arXiv preprint arXiv:1709.04893, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32.5,
+ "bbox_fs": [
+ 106,
+ 543,
+ 505,
+ 567
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 572,
+ 470,
+ 585
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 471,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 471,
+ 586
+ ],
+ "score": 1.0,
+ "content": "Taco S Cohen and Max Welling. Steerable cnns. arXiv preprint arXiv:1612.08498, 2016b.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 34,
+ "bbox_fs": [
+ 105,
+ 572,
+ 471,
+ 586
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 591,
+ 502,
+ 614
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 591,
+ 504,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 591,
+ 504,
+ 604
+ ],
+ "score": 1.0,
+ "content": "Taco S Cohen, Maurice Weiler, Berkay Kicanaoglu, and Max Welling. Gauge equivariant convolu-",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 116,
+ 602,
+ 443,
+ 615
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 602,
+ 443,
+ 615
+ ],
+ "score": 1.0,
+ "content": "tional networks and the icosahedral cnn. arXiv preprint arXiv:1902.04615, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35.5,
+ "bbox_fs": [
+ 106,
+ 591,
+ 504,
+ 615
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 621,
+ 503,
+ 644
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 621,
+ 505,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 621,
+ 505,
+ 633
+ ],
+ "score": 1.0,
+ "content": "Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 116,
+ 632,
+ 329,
+ 643
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 632,
+ 329,
+ 643
+ ],
+ "score": 1.0,
+ "content": "with cutout. arXiv preprint arXiv:1708.04552, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 37.5,
+ "bbox_fs": [
+ 106,
+ 621,
+ 505,
+ 643
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 650,
+ 505,
+ 684
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 505,
+ 664
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 505,
+ 664
+ ],
+ "score": 1.0,
+ "content": "Carlos Esteves, Christine Allen-Blanchette, Ameesh Makadia, and Kostas Daniilidis. Learning so",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 116,
+ 661,
+ 505,
+ 675
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 661,
+ 505,
+ 675
+ ],
+ "score": 1.0,
+ "content": "(3) equivariant representations with spherical cnns. In Proceedings of the European Conference",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 116,
+ 672,
+ 305,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 672,
+ 305,
+ 685
+ ],
+ "score": 1.0,
+ "content": "on Computer Vision (ECCV), pp. 52–68, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40,
+ "bbox_fs": [
+ 105,
+ 649,
+ 505,
+ 685
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 690,
+ 471,
+ 703
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 690,
+ 471,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 690,
+ 471,
+ 703
+ ],
+ "score": 1.0,
+ "content": "Gerald B Folland. A course in abstract harmonic analysis. Chapman and Hall/CRC, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 42,
+ "bbox_fs": [
+ 106,
+ 690,
+ 471,
+ 703
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 709,
+ 502,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 708,
+ 504,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 708,
+ 504,
+ 722
+ ],
+ "score": 1.0,
+ "content": "William T. Freeman and Edward H Adelson. The design and use of steerable filters. IEEE Transac-",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 115,
+ 721,
+ 398,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 721,
+ 398,
+ 732
+ ],
+ "score": 1.0,
+ "content": "tions on Pattern Analysis & Machine Intelligence, (9):891–906, 1991.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43.5,
+ "bbox_fs": [
+ 106,
+ 708,
+ 504,
+ 732
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 82,
+ 504,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "Rohan Ghosh and Anupam K Gupta. Scale steerable filters for locally scale-invariant convolutional",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 115,
+ 93,
+ 349,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 93,
+ 349,
+ 106
+ ],
+ "score": 1.0,
+ "content": "neural networks. arXiv preprint arXiv:1906.03861, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 111,
+ 503,
+ 146
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 111,
+ 505,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 111,
+ 505,
+ 126
+ ],
+ "score": 1.0,
+ "content": "Joao F Henriques and Andrea Vedaldi. Warped convolutions: Efficient invariance to spatial trans-",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 115,
+ 123,
+ 505,
+ 136
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 123,
+ 505,
+ 136
+ ],
+ "score": 1.0,
+ "content": "formations. In Proceedings of the 34th International Conference on Machine Learning-Volume",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 116,
+ 133,
+ 270,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 133,
+ 270,
+ 147
+ ],
+ "score": 1.0,
+ "content": "70, pp. 1461–1469. JMLR. org, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 152,
+ 504,
+ 176
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 151,
+ 506,
+ 166
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 151,
+ 506,
+ 166
+ ],
+ "score": 1.0,
+ "content": "Emiel Hoogeboom, Jorn WT Peters, Taco S Cohen, and Max Welling. Hexaconv. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 116,
+ 163,
+ 219,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 163,
+ 219,
+ 174
+ ],
+ "score": 1.0,
+ "content": "arXiv:1803.02108, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 182,
+ 504,
+ 205
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 182,
+ 505,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 182,
+ 505,
+ 195
+ ],
+ "score": 1.0,
+ "content": "Taizo Iijima. Basic theory of pattern observation. Technical Group on Automata and Automatic",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 116,
+ 192,
+ 217,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 192,
+ 217,
+ 205
+ ],
+ "score": 1.0,
+ "content": "Control, pp. 3–32, 1959.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 212,
+ 504,
+ 235
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 505,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 505,
+ 225
+ ],
+ "score": 1.0,
+ "content": "Jorn-Henrik Jacobsen, Bert De Brabandere, and Arnold WM Smeulders. Dynamic steerable blocks ¨",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 116,
+ 222,
+ 387,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 222,
+ 387,
+ 235
+ ],
+ "score": 1.0,
+ "content": "in deep residual networks. arXiv preprint arXiv:1706.00598, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 9.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 241,
+ 505,
+ 265
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 241,
+ 506,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 241,
+ 506,
+ 255
+ ],
+ "score": 1.0,
+ "content": "Angjoo Kanazawa, Abhishek Sharma, and David Jacobs. Locally scale-invariant convolutional",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 115,
+ 253,
+ 344,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 253,
+ 344,
+ 265
+ ],
+ "score": 1.0,
+ "content": "neural networks. arXiv preprint arXiv:1412.5104, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 271,
+ 505,
+ 294
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 505,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 505,
+ 285
+ ],
+ "score": 1.0,
+ "content": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 115,
+ 282,
+ 214,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 282,
+ 214,
+ 294
+ ],
+ "score": 1.0,
+ "content": "arXiv:1412.6980, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 300,
+ 466,
+ 313
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 466,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 466,
+ 315
+ ],
+ "score": 1.0,
+ "content": "Jan J Koenderink. The structure of images. Biological cybernetics, 50(5):363–370, 1984.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 320,
+ 504,
+ 343
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 318,
+ 506,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 318,
+ 506,
+ 334
+ ],
+ "score": 1.0,
+ "content": "Risi Kondor. N-body networks: a covariant hierarchical neural network architecture for learning",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 331,
+ 353,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 331,
+ 353,
+ 343
+ ],
+ "score": 1.0,
+ "content": "atomic potentials. arXiv preprint arXiv:1803.01588, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 16.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 349,
+ 504,
+ 373
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 349,
+ 504,
+ 362
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 349,
+ 504,
+ 362
+ ],
+ "score": 1.0,
+ "content": "Risi Kondor and Shubhendu Trivedi. On the generalization of equivariance and convolution in neural",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 115,
+ 361,
+ 449,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 361,
+ 449,
+ 373
+ ],
+ "score": 1.0,
+ "content": "networks to the action of compact groups. arXiv preprint arXiv:1802.03690, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 379,
+ 504,
+ 403
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 379,
+ 505,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 379,
+ 505,
+ 392
+ ],
+ "score": 1.0,
+ "content": "Yann LeCun, Leon Bottou, Yoshua Bengio, Patrick Haffner, et al. Gradient-based learning applied ´",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 115,
+ 391,
+ 427,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 391,
+ 427,
+ 403
+ ],
+ "score": 1.0,
+ "content": "to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 408,
+ 504,
+ 432
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 505,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 505,
+ 422
+ ],
+ "score": 1.0,
+ "content": "Tony Lindeberg. Scale-space theory in computer vision, volume 256. Springer Science & Business",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 115,
+ 420,
+ 173,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 420,
+ 173,
+ 432
+ ],
+ "score": 1.0,
+ "content": "Media, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 438,
+ 504,
+ 462
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 437,
+ 506,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 437,
+ 506,
+ 453
+ ],
+ "score": 1.0,
+ "content": "Diego Marcos, Benjamin Kellenberger, Sylvain Lobry, and Devis Tuia. Scale equivariance in cnns",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 116,
+ 450,
+ 353,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 450,
+ 353,
+ 462
+ ],
+ "score": 1.0,
+ "content": "with vector fields. arXiv preprint arXiv:1807.11783, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 468,
+ 503,
+ 491
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 482
+ ],
+ "score": 1.0,
+ "content": "Muthuvel Murugan, KV Subrahmanyam, and Siruseri IT Park. So (2)-equivariance in neural net-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 116,
+ 479,
+ 246,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 479,
+ 246,
+ 493
+ ],
+ "score": 1.0,
+ "content": "works using tensor nonlinearity.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 498,
+ 504,
+ 522
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 498,
+ 505,
+ 510
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 498,
+ 505,
+ 510
+ ],
+ "score": 1.0,
+ "content": "Pietro Perona and Jitendra Malik. Scale-space and edge detection using anisotropic diffusion. IEEE",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 117,
+ 509,
+ 443,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 509,
+ 443,
+ 522
+ ],
+ "score": 1.0,
+ "content": "Transactions on pattern analysis and machine intelligence, 12(7):629–639, 1990.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 527,
+ 504,
+ 551
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": "Kai Sheng Tai, Peter Bailis, and Gregory Valiant. Equivariant transformer networks. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 116,
+ 539,
+ 219,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 539,
+ 219,
+ 550
+ ],
+ "score": 1.0,
+ "content": "arXiv:1901.11399, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 557,
+ 504,
+ 591
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 570
+ ],
+ "score": 1.0,
+ "content": "Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 568,
+ 505,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 568,
+ 505,
+ 581
+ ],
+ "score": 1.0,
+ "content": "Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 116,
+ 579,
+ 311,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 579,
+ 311,
+ 592
+ ],
+ "score": 1.0,
+ "content": "clouds. arXiv preprint arXiv:1802.08219, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 597,
+ 504,
+ 621
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 506,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 506,
+ 612
+ ],
+ "score": 1.0,
+ "content": "Matej Ulicny, Vladimir A Krylov, and Rozenn Dahyot. Harmonic networks with limited training",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 116,
+ 610,
+ 317,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 610,
+ 317,
+ 621
+ ],
+ "score": 1.0,
+ "content": "samples. arXiv preprint arXiv:1905.00135, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 651
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "Maurice Weiler and Gabriele Cesa. General e (2)-equivariant steerable cnns. In Advances in Neural",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 116,
+ 638,
+ 349,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 638,
+ 349,
+ 651
+ ],
+ "score": 1.0,
+ "content": "Information Processing Systems, pp. 14334–14345, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 37.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 657,
+ 504,
+ 691
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 657,
+ 505,
+ 671
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 657,
+ 505,
+ 671
+ ],
+ "score": 1.0,
+ "content": "Maurice Weiler, Mario Geiger, Max Welling, Wouter Boomsma, and Taco Cohen. 3d steerable cnns:",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 115,
+ 669,
+ 505,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 669,
+ 505,
+ 681
+ ],
+ "score": 1.0,
+ "content": "Learning rotationally equivariant features in volumetric data. In Advances in Neural Information",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 115,
+ 679,
+ 303,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 679,
+ 303,
+ 693
+ ],
+ "score": 1.0,
+ "content": "Processing Systems, pp. 10381–10392, 2018a.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 699,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "Maurice Weiler, Fred A Hamprecht, and Martin Storath. Learning steerable filters for rotation equiv-",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 116,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "ariant cnns. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 115,
+ 721,
+ 201,
+ 731
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 721,
+ 201,
+ 731
+ ],
+ "score": 1.0,
+ "content": "pp. 849–858, 2018b.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "page_idx": 9,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 294,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 764
+ ],
+ "score": 1.0,
+ "content": "10",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 82,
+ 504,
+ 105
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "Rohan Ghosh and Anupam K Gupta. Scale steerable filters for locally scale-invariant convolutional",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 115,
+ 93,
+ 349,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 93,
+ 349,
+ 106
+ ],
+ "score": 1.0,
+ "content": "neural networks. arXiv preprint arXiv:1906.03861, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5,
+ "bbox_fs": [
+ 106,
+ 82,
+ 505,
+ 106
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 111,
+ 503,
+ 146
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 111,
+ 505,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 111,
+ 505,
+ 126
+ ],
+ "score": 1.0,
+ "content": "Joao F Henriques and Andrea Vedaldi. Warped convolutions: Efficient invariance to spatial trans-",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 115,
+ 123,
+ 505,
+ 136
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 123,
+ 505,
+ 136
+ ],
+ "score": 1.0,
+ "content": "formations. In Proceedings of the 34th International Conference on Machine Learning-Volume",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 116,
+ 133,
+ 270,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 133,
+ 270,
+ 147
+ ],
+ "score": 1.0,
+ "content": "70, pp. 1461–1469. JMLR. org, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3,
+ "bbox_fs": [
+ 105,
+ 111,
+ 505,
+ 147
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 152,
+ 504,
+ 176
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 151,
+ 506,
+ 166
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 151,
+ 506,
+ 166
+ ],
+ "score": 1.0,
+ "content": "Emiel Hoogeboom, Jorn WT Peters, Taco S Cohen, and Max Welling. Hexaconv. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 116,
+ 163,
+ 219,
+ 174
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 163,
+ 219,
+ 174
+ ],
+ "score": 1.0,
+ "content": "arXiv:1803.02108, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5,
+ "bbox_fs": [
+ 105,
+ 151,
+ 506,
+ 174
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 182,
+ 504,
+ 205
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 182,
+ 505,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 182,
+ 505,
+ 195
+ ],
+ "score": 1.0,
+ "content": "Taizo Iijima. Basic theory of pattern observation. Technical Group on Automata and Automatic",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 116,
+ 192,
+ 217,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 192,
+ 217,
+ 205
+ ],
+ "score": 1.0,
+ "content": "Control, pp. 3–32, 1959.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5,
+ "bbox_fs": [
+ 106,
+ 182,
+ 505,
+ 205
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 212,
+ 504,
+ 235
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 505,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 505,
+ 225
+ ],
+ "score": 1.0,
+ "content": "Jorn-Henrik Jacobsen, Bert De Brabandere, and Arnold WM Smeulders. Dynamic steerable blocks ¨",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 116,
+ 222,
+ 387,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 222,
+ 387,
+ 235
+ ],
+ "score": 1.0,
+ "content": "in deep residual networks. arXiv preprint arXiv:1706.00598, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 9.5,
+ "bbox_fs": [
+ 105,
+ 211,
+ 505,
+ 235
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 241,
+ 505,
+ 265
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 241,
+ 506,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 241,
+ 506,
+ 255
+ ],
+ "score": 1.0,
+ "content": "Angjoo Kanazawa, Abhishek Sharma, and David Jacobs. Locally scale-invariant convolutional",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 115,
+ 253,
+ 344,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 253,
+ 344,
+ 265
+ ],
+ "score": 1.0,
+ "content": "neural networks. arXiv preprint arXiv:1412.5104, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11.5,
+ "bbox_fs": [
+ 105,
+ 241,
+ 506,
+ 265
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 271,
+ 505,
+ 294
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 505,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 270,
+ 505,
+ 285
+ ],
+ "score": 1.0,
+ "content": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 115,
+ 282,
+ 214,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 282,
+ 214,
+ 294
+ ],
+ "score": 1.0,
+ "content": "arXiv:1412.6980, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5,
+ "bbox_fs": [
+ 105,
+ 270,
+ 505,
+ 294
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 300,
+ 466,
+ 313
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 466,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 466,
+ 315
+ ],
+ "score": 1.0,
+ "content": "Jan J Koenderink. The structure of images. Biological cybernetics, 50(5):363–370, 1984.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 15,
+ "bbox_fs": [
+ 105,
+ 300,
+ 466,
+ 315
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 320,
+ 504,
+ 343
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 318,
+ 506,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 318,
+ 506,
+ 334
+ ],
+ "score": 1.0,
+ "content": "Risi Kondor. N-body networks: a covariant hierarchical neural network architecture for learning",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 331,
+ 353,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 331,
+ 353,
+ 343
+ ],
+ "score": 1.0,
+ "content": "atomic potentials. arXiv preprint arXiv:1803.01588, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 16.5,
+ "bbox_fs": [
+ 105,
+ 318,
+ 506,
+ 343
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 349,
+ 504,
+ 373
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 349,
+ 504,
+ 362
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 349,
+ 504,
+ 362
+ ],
+ "score": 1.0,
+ "content": "Risi Kondor and Shubhendu Trivedi. On the generalization of equivariance and convolution in neural",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 115,
+ 361,
+ 449,
+ 373
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 361,
+ 449,
+ 373
+ ],
+ "score": 1.0,
+ "content": "networks to the action of compact groups. arXiv preprint arXiv:1802.03690, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18.5,
+ "bbox_fs": [
+ 106,
+ 349,
+ 504,
+ 373
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 379,
+ 504,
+ 403
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 379,
+ 505,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 379,
+ 505,
+ 392
+ ],
+ "score": 1.0,
+ "content": "Yann LeCun, Leon Bottou, Yoshua Bengio, Patrick Haffner, et al. Gradient-based learning applied ´",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 115,
+ 391,
+ 427,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 391,
+ 427,
+ 403
+ ],
+ "score": 1.0,
+ "content": "to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20.5,
+ "bbox_fs": [
+ 106,
+ 379,
+ 505,
+ 403
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 408,
+ 504,
+ 432
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 505,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 408,
+ 505,
+ 422
+ ],
+ "score": 1.0,
+ "content": "Tony Lindeberg. Scale-space theory in computer vision, volume 256. Springer Science & Business",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 115,
+ 420,
+ 173,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 420,
+ 173,
+ 432
+ ],
+ "score": 1.0,
+ "content": "Media, 2013.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.5,
+ "bbox_fs": [
+ 105,
+ 408,
+ 505,
+ 432
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 438,
+ 504,
+ 462
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 437,
+ 506,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 437,
+ 506,
+ 453
+ ],
+ "score": 1.0,
+ "content": "Diego Marcos, Benjamin Kellenberger, Sylvain Lobry, and Devis Tuia. Scale equivariance in cnns",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 116,
+ 450,
+ 353,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 450,
+ 353,
+ 462
+ ],
+ "score": 1.0,
+ "content": "with vector fields. arXiv preprint arXiv:1807.11783, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5,
+ "bbox_fs": [
+ 105,
+ 437,
+ 506,
+ 462
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 468,
+ 503,
+ 491
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 482
+ ],
+ "score": 1.0,
+ "content": "Muthuvel Murugan, KV Subrahmanyam, and Siruseri IT Park. So (2)-equivariance in neural net-",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 116,
+ 479,
+ 246,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 479,
+ 246,
+ 493
+ ],
+ "score": 1.0,
+ "content": "works using tensor nonlinearity.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 105,
+ 468,
+ 505,
+ 493
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 498,
+ 504,
+ 522
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 498,
+ 505,
+ 510
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 498,
+ 505,
+ 510
+ ],
+ "score": 1.0,
+ "content": "Pietro Perona and Jitendra Malik. Scale-space and edge detection using anisotropic diffusion. IEEE",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 117,
+ 509,
+ 443,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 509,
+ 443,
+ 522
+ ],
+ "score": 1.0,
+ "content": "Transactions on pattern analysis and machine intelligence, 12(7):629–639, 1990.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28.5,
+ "bbox_fs": [
+ 106,
+ 498,
+ 505,
+ 522
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 527,
+ 504,
+ 551
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": "Kai Sheng Tai, Peter Bailis, and Gregory Valiant. Equivariant transformer networks. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 116,
+ 539,
+ 219,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 539,
+ 219,
+ 550
+ ],
+ "score": 1.0,
+ "content": "arXiv:1901.11399, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.5,
+ "bbox_fs": [
+ 105,
+ 527,
+ 505,
+ 550
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 557,
+ 504,
+ 591
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 558,
+ 505,
+ 570
+ ],
+ "score": 1.0,
+ "content": "Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 568,
+ 505,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 568,
+ 505,
+ 581
+ ],
+ "score": 1.0,
+ "content": "Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 116,
+ 579,
+ 311,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 579,
+ 311,
+ 592
+ ],
+ "score": 1.0,
+ "content": "clouds. arXiv preprint arXiv:1802.08219, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 33,
+ "bbox_fs": [
+ 105,
+ 558,
+ 505,
+ 592
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 597,
+ 504,
+ 621
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 506,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 506,
+ 612
+ ],
+ "score": 1.0,
+ "content": "Matej Ulicny, Vladimir A Krylov, and Rozenn Dahyot. Harmonic networks with limited training",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 116,
+ 610,
+ 317,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 610,
+ 317,
+ 621
+ ],
+ "score": 1.0,
+ "content": "samples. arXiv preprint arXiv:1905.00135, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35.5,
+ "bbox_fs": [
+ 105,
+ 596,
+ 506,
+ 621
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 651
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "Maurice Weiler and Gabriele Cesa. General e (2)-equivariant steerable cnns. In Advances in Neural",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 116,
+ 638,
+ 349,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 638,
+ 349,
+ 651
+ ],
+ "score": 1.0,
+ "content": "Information Processing Systems, pp. 14334–14345, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 37.5,
+ "bbox_fs": [
+ 106,
+ 627,
+ 505,
+ 651
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 657,
+ 504,
+ 691
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 657,
+ 505,
+ 671
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 657,
+ 505,
+ 671
+ ],
+ "score": 1.0,
+ "content": "Maurice Weiler, Mario Geiger, Max Welling, Wouter Boomsma, and Taco Cohen. 3d steerable cnns:",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 115,
+ 669,
+ 505,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 669,
+ 505,
+ 681
+ ],
+ "score": 1.0,
+ "content": "Learning rotationally equivariant features in volumetric data. In Advances in Neural Information",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 115,
+ 679,
+ 303,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 679,
+ 303,
+ 693
+ ],
+ "score": 1.0,
+ "content": "Processing Systems, pp. 10381–10392, 2018a.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40,
+ "bbox_fs": [
+ 106,
+ 657,
+ 505,
+ 693
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 699,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "Maurice Weiler, Fred A Hamprecht, and Martin Storath. Learning steerable filters for rotation equiv-",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 116,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "ariant cnns. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 115,
+ 721,
+ 201,
+ 731
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 721,
+ 201,
+ 731
+ ],
+ "score": 1.0,
+ "content": "pp. 849–858, 2018b.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43,
+ "bbox_fs": [
+ 105,
+ 698,
+ 505,
+ 731
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 81,
+ 506,
+ 266
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "Andrew P Witkin. Scale-space filtering. In Readings in Computer Vision, pp. 329–332. Elsevier,",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 116,
+ 94,
+ 142,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 94,
+ 142,
+ 105
+ ],
+ "score": 1.0,
+ "content": "1987.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 113,
+ 505,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 113,
+ 505,
+ 124
+ ],
+ "score": 1.0,
+ "content": "Daniel Worrall and Gabriel Brostow. Cubenet: Equivariance to 3d rotation and translation. In",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 116,
+ 124,
+ 482,
+ 136
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 124,
+ 482,
+ 136
+ ],
+ "score": 1.0,
+ "content": "Proceedings of the European Conference on Computer Vision (ECCV), pp. 567–584, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 141,
+ 506,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 141,
+ 506,
+ 156
+ ],
+ "score": 1.0,
+ "content": "Daniel E Worrall and Max Welling. Deep scale-spaces: Equivariance over scale. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 116,
+ 154,
+ 219,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 154,
+ 219,
+ 165
+ ],
+ "score": 1.0,
+ "content": "arXiv:1905.11697, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 172,
+ 505,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 172,
+ 505,
+ 185
+ ],
+ "score": 1.0,
+ "content": "Daniel E Worrall, Stephan J Garbin, Daniyar Turmukhambetov, and Gabriel J Brostow. Harmonic",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 116,
+ 184,
+ 505,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 184,
+ 505,
+ 197
+ ],
+ "score": 1.0,
+ "content": "networks: Deep translation and rotation equivariance. In Proceedings of the IEEE Conference on",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 116,
+ 194,
+ 379,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 194,
+ 379,
+ 208
+ ],
+ "score": 1.0,
+ "content": "Computer Vision and Pattern Recognition, pp. 5028–5037, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 212,
+ 505,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 212,
+ 505,
+ 228
+ ],
+ "score": 1.0,
+ "content": "Yichong Xu, Tianjun Xiao, Jiaxing Zhang, Kuiyuan Yang, and Zheng Zhang. Scale-invariant con-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 116,
+ 224,
+ 387,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 224,
+ 387,
+ 238
+ ],
+ "score": 1.0,
+ "content": "volutional neural networks. arXiv preprint arXiv:1411.6369, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 243,
+ 505,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 243,
+ 297,
+ 256
+ ],
+ "score": 1.0,
+ "content": "Sergey Zagoruyko and Nikos Komodakis.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 313,
+ 244,
+ 423,
+ 255
+ ],
+ "score": 1.0,
+ "content": "Wide residual networks.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 440,
+ 244,
+ 505,
+ 256
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 115,
+ 254,
+ 219,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 254,
+ 219,
+ 266
+ ],
+ "score": 1.0,
+ "content": "arXiv:1605.07146, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "page_idx": 10,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 27,
+ 294,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 293,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2020",
+ "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": "list",
+ "bbox": [
+ 105,
+ 81,
+ 506,
+ 266
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 81,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "Andrew P Witkin. Scale-space filtering. In Readings in Computer Vision, pp. 329–332. Elsevier,",
+ "type": "text"
+ }
+ ],
+ "index": 0,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 116,
+ 94,
+ 142,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 94,
+ 142,
+ 105
+ ],
+ "score": 1.0,
+ "content": "1987.",
+ "type": "text"
+ }
+ ],
+ "index": 1,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 113,
+ 505,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 113,
+ 505,
+ 124
+ ],
+ "score": 1.0,
+ "content": "Daniel Worrall and Gabriel Brostow. Cubenet: Equivariance to 3d rotation and translation. In",
+ "type": "text"
+ }
+ ],
+ "index": 2,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 116,
+ 124,
+ 482,
+ 136
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 124,
+ 482,
+ 136
+ ],
+ "score": 1.0,
+ "content": "Proceedings of the European Conference on Computer Vision (ECCV), pp. 567–584, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 3,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 141,
+ 506,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 141,
+ 506,
+ 156
+ ],
+ "score": 1.0,
+ "content": "Daniel E Worrall and Max Welling. Deep scale-spaces: Equivariance over scale. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 4,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 116,
+ 154,
+ 219,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 154,
+ 219,
+ 165
+ ],
+ "score": 1.0,
+ "content": "arXiv:1905.11697, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 5,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 172,
+ 505,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 172,
+ 505,
+ 185
+ ],
+ "score": 1.0,
+ "content": "Daniel E Worrall, Stephan J Garbin, Daniyar Turmukhambetov, and Gabriel J Brostow. Harmonic",
+ "type": "text"
+ }
+ ],
+ "index": 6,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 116,
+ 184,
+ 505,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 184,
+ 505,
+ 197
+ ],
+ "score": 1.0,
+ "content": "networks: Deep translation and rotation equivariance. In Proceedings of the IEEE Conference on",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 116,
+ 194,
+ 379,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 194,
+ 379,
+ 208
+ ],
+ "score": 1.0,
+ "content": "Computer Vision and Pattern Recognition, pp. 5028–5037, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 8,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 105,
+ 212,
+ 505,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 212,
+ 505,
+ 228
+ ],
+ "score": 1.0,
+ "content": "Yichong Xu, Tianjun Xiao, Jiaxing Zhang, Kuiyuan Yang, and Zheng Zhang. Scale-invariant con-",
+ "type": "text"
+ }
+ ],
+ "index": 9,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 116,
+ 224,
+ 387,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 224,
+ 387,
+ 238
+ ],
+ "score": 1.0,
+ "content": "volutional neural networks. arXiv preprint arXiv:1411.6369, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 10,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 106,
+ 243,
+ 505,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 243,
+ 297,
+ 256
+ ],
+ "score": 1.0,
+ "content": "Sergey Zagoruyko and Nikos Komodakis.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 313,
+ 244,
+ 423,
+ 255
+ ],
+ "score": 1.0,
+ "content": "Wide residual networks.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 440,
+ 244,
+ 505,
+ 256
+ ],
+ "score": 1.0,
+ "content": "arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 11,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 115,
+ 254,
+ 219,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 254,
+ 219,
+ 266
+ ],
+ "score": 1.0,
+ "content": "arXiv:1605.07146, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 12,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 105,
+ 81,
+ 506,
+ 266
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 81,
+ 263,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 79,
+ 264,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 79,
+ 264,
+ 97
+ ],
+ "score": 1.0,
+ "content": "A PROOF OF EQUIVARIANCE",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 103,
+ 105,
+ 469,
+ 119
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 469,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 469,
+ 119
+ ],
+ "score": 1.0,
+ "content": "Let us first show that scale-convolution defined in Equation 6 is equivariant to translations.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 185,
+ 137,
+ 425,
+ 237
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 185,
+ 137,
+ 425,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 185,
+ 137,
+ 425,
+ 237
+ ],
+ "score": 0.95,
+ "content": "\\begin{array} { l } { [ L _ { \\hat { t } } [ f ] \\star _ { H } \\psi _ { \\sigma } ] ( s , t ) = \\displaystyle \\sum _ { s ^ { \\prime } } [ L _ { \\hat { t } } [ f ] ( s ^ { \\prime } , \\cdot ) \\star \\psi _ { s \\sigma } ( s ^ { - 1 } s ^ { \\prime } , \\cdot ) ] ( t ) } \\\\ { = \\displaystyle \\sum _ { s ^ { \\prime } } L _ { \\hat { t } } [ f ( s ^ { \\prime } , \\cdot ) \\star \\psi _ { s \\sigma } ( s ^ { - 1 } s ^ { \\prime } , \\cdot ) ] ( t ) } \\\\ { = L _ { \\hat { t } } \\Big \\{ \\displaystyle \\sum _ { s ^ { \\prime } } [ f ( s ^ { \\prime } , \\cdot ) \\star \\psi _ { s \\sigma } ( s ^ { - 1 } s ^ { \\prime } , \\cdot ) ] \\Big \\} ( t ) } \\\\ { = L _ { \\hat { t } } [ f \\star _ { H } \\psi _ { \\sigma } ] ( s , t ) } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "a6d67739f8ff7e9b634adeaec82fac0aae10937b270cb5821f7c2f2af3c98af3.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 185,
+ 137,
+ 425,
+ 153.66666666666666
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 185,
+ 153.66666666666666,
+ 425,
+ 170.33333333333331
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 185,
+ 170.33333333333331,
+ 425,
+ 186.99999999999997
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 185,
+ 186.99999999999997,
+ 425,
+ 203.66666666666663
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 185,
+ 203.66666666666663,
+ 425,
+ 220.3333333333333
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 185,
+ 220.3333333333333,
+ 425,
+ 236.99999999999994
+ ],
+ "spans": [],
+ "index": 7
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 245,
+ 411,
+ 258
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 245,
+ 411,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 245,
+ 411,
+ 258
+ ],
+ "score": 1.0,
+ "content": "Now we show that scale convolution is equivariant to scale transformations:",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 176,
+ 263,
+ 434,
+ 374
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 176,
+ 263,
+ 434,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 176,
+ 263,
+ 434,
+ 374
+ ],
+ "score": 0.95,
+ "content": "\\begin{array} { l } { [ L _ { \\delta } [ f ] \\star _ { H } \\psi _ { \\sigma } ] ( s , t ) = \\displaystyle \\sum _ { s ^ { \\prime } } [ L _ { \\delta } [ f ] ( s ^ { \\prime } , \\cdot ) \\star \\psi _ { s \\sigma } ( s ^ { - 1 } s ^ { \\prime } , \\cdot ) ] ( t ) } \\\\ { = \\displaystyle \\sum _ { s ^ { \\prime } } L _ { \\delta } [ f ( \\hat { s } ^ { - 1 } s ^ { \\prime } , \\cdot ) \\star \\psi _ { \\hat { s } ^ { - 1 } s \\sigma } ( s ^ { - 1 } s ^ { \\prime } , \\cdot ) ] ( t ) } \\\\ { = \\displaystyle \\sum _ { s ^ { \\prime \\prime } } [ f ( s ^ { \\prime \\prime } , \\cdot ) \\star \\psi _ { \\hat { s } ^ { - 1 } s \\sigma } ( \\hat { s } s ^ { - 1 } s ^ { \\prime \\prime } , \\cdot ) ] ( \\hat { s } ^ { - 1 } t ) } \\\\ { = [ f \\star _ { H } \\psi _ { \\sigma } ] ( \\hat { s } ^ { - 1 } s , \\hat { s } ^ { - 1 } t ) } \\\\ { = L _ { \\delta } [ f \\star _ { H } \\psi _ { \\sigma } ] ( s , t ) } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "a10c6b05a353e59e312fdd7276c15886a1566d7cd04c9cade1a944a14b3949e3.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 176,
+ 263,
+ 434,
+ 300.0
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 176,
+ 300.0,
+ 434,
+ 337.0
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 176,
+ 337.0,
+ 434,
+ 374.0
+ ],
+ "spans": [],
+ "index": 11
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 384,
+ 365,
+ 397
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 365,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 365,
+ 399
+ ],
+ "score": 1.0,
+ "content": "Finally, we can use the property of semidirect product of groups",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 119,
+ 402,
+ 475,
+ 416
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 402,
+ 475,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 402,
+ 475,
+ 416
+ ],
+ "score": 0.89,
+ "content": "L _ { \\hat { s t } } [ f ] \\star _ { H } \\psi _ { \\sigma } = L _ { \\hat { s } } L _ { \\hat { t } } [ f ] \\star _ { H } \\psi _ { \\sigma } = L _ { \\hat { s } } [ L _ { \\hat { t } } [ f ] \\star _ { H } \\psi _ { \\sigma } ] = L _ { \\hat { s } } L _ { \\hat { t } } [ f \\star _ { H } \\psi _ { \\sigma } ] = L _ { \\hat { s } \\hat { t } } [ f \\star _ { H } \\psi _ { \\sigma } ] = L _ { \\hat { s t } } [ f \\star _ { H } \\psi _ { \\sigma } ]",
+ "type": "interline_equation",
+ "image_path": "209a5ce9ce9f2fb60134f9570f24b4b4f44e3552f47a5c97a83498b4435f0da6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 119,
+ 402,
+ 475,
+ 416
+ ],
+ "spans": [],
+ "index": 13
+ }
+ ]
+ }
+ ],
+ "page_idx": 11,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 765
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 765
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 15,
+ "width": 14
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 81,
+ 263,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 79,
+ 264,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 79,
+ 264,
+ 97
+ ],
+ "score": 1.0,
+ "content": "A PROOF OF EQUIVARIANCE",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 103,
+ 105,
+ 469,
+ 119
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 469,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 105,
+ 469,
+ 119
+ ],
+ "score": 1.0,
+ "content": "Let us first show that scale-convolution defined in Equation 6 is equivariant to translations.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 1,
+ "bbox_fs": [
+ 105,
+ 105,
+ 469,
+ 119
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 185,
+ 137,
+ 425,
+ 237
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 185,
+ 137,
+ 425,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 185,
+ 137,
+ 425,
+ 237
+ ],
+ "score": 0.95,
+ "content": "\\begin{array} { l } { [ L _ { \\hat { t } } [ f ] \\star _ { H } \\psi _ { \\sigma } ] ( s , t ) = \\displaystyle \\sum _ { s ^ { \\prime } } [ L _ { \\hat { t } } [ f ] ( s ^ { \\prime } , \\cdot ) \\star \\psi _ { s \\sigma } ( s ^ { - 1 } s ^ { \\prime } , \\cdot ) ] ( t ) } \\\\ { = \\displaystyle \\sum _ { s ^ { \\prime } } L _ { \\hat { t } } [ f ( s ^ { \\prime } , \\cdot ) \\star \\psi _ { s \\sigma } ( s ^ { - 1 } s ^ { \\prime } , \\cdot ) ] ( t ) } \\\\ { = L _ { \\hat { t } } \\Big \\{ \\displaystyle \\sum _ { s ^ { \\prime } } [ f ( s ^ { \\prime } , \\cdot ) \\star \\psi _ { s \\sigma } ( s ^ { - 1 } s ^ { \\prime } , \\cdot ) ] \\Big \\} ( t ) } \\\\ { = L _ { \\hat { t } } [ f \\star _ { H } \\psi _ { \\sigma } ] ( s , t ) } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "a6d67739f8ff7e9b634adeaec82fac0aae10937b270cb5821f7c2f2af3c98af3.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 185,
+ 137,
+ 425,
+ 153.66666666666666
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 185,
+ 153.66666666666666,
+ 425,
+ 170.33333333333331
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 185,
+ 170.33333333333331,
+ 425,
+ 186.99999999999997
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 185,
+ 186.99999999999997,
+ 425,
+ 203.66666666666663
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 185,
+ 203.66666666666663,
+ 425,
+ 220.3333333333333
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 185,
+ 220.3333333333333,
+ 425,
+ 236.99999999999994
+ ],
+ "spans": [],
+ "index": 7
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 245,
+ 411,
+ 258
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 245,
+ 411,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 245,
+ 411,
+ 258
+ ],
+ "score": 1.0,
+ "content": "Now we show that scale convolution is equivariant to scale transformations:",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 8,
+ "bbox_fs": [
+ 106,
+ 245,
+ 411,
+ 258
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 176,
+ 263,
+ 434,
+ 374
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 176,
+ 263,
+ 434,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 176,
+ 263,
+ 434,
+ 374
+ ],
+ "score": 0.95,
+ "content": "\\begin{array} { l } { [ L _ { \\delta } [ f ] \\star _ { H } \\psi _ { \\sigma } ] ( s , t ) = \\displaystyle \\sum _ { s ^ { \\prime } } [ L _ { \\delta } [ f ] ( s ^ { \\prime } , \\cdot ) \\star \\psi _ { s \\sigma } ( s ^ { - 1 } s ^ { \\prime } , \\cdot ) ] ( t ) } \\\\ { = \\displaystyle \\sum _ { s ^ { \\prime } } L _ { \\delta } [ f ( \\hat { s } ^ { - 1 } s ^ { \\prime } , \\cdot ) \\star \\psi _ { \\hat { s } ^ { - 1 } s \\sigma } ( s ^ { - 1 } s ^ { \\prime } , \\cdot ) ] ( t ) } \\\\ { = \\displaystyle \\sum _ { s ^ { \\prime \\prime } } [ f ( s ^ { \\prime \\prime } , \\cdot ) \\star \\psi _ { \\hat { s } ^ { - 1 } s \\sigma } ( \\hat { s } s ^ { - 1 } s ^ { \\prime \\prime } , \\cdot ) ] ( \\hat { s } ^ { - 1 } t ) } \\\\ { = [ f \\star _ { H } \\psi _ { \\sigma } ] ( \\hat { s } ^ { - 1 } s , \\hat { s } ^ { - 1 } t ) } \\\\ { = L _ { \\delta } [ f \\star _ { H } \\psi _ { \\sigma } ] ( s , t ) } \\end{array}",
+ "type": "interline_equation",
+ "image_path": "a10c6b05a353e59e312fdd7276c15886a1566d7cd04c9cade1a944a14b3949e3.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 176,
+ 263,
+ 434,
+ 300.0
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 176,
+ 300.0,
+ 434,
+ 337.0
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 176,
+ 337.0,
+ 434,
+ 374.0
+ ],
+ "spans": [],
+ "index": 11
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 384,
+ 365,
+ 397
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 365,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 365,
+ 399
+ ],
+ "score": 1.0,
+ "content": "Finally, we can use the property of semidirect product of groups",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 12,
+ "bbox_fs": [
+ 105,
+ 383,
+ 365,
+ 399
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 119,
+ 402,
+ 475,
+ 416
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 402,
+ 475,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 402,
+ 475,
+ 416
+ ],
+ "score": 0.89,
+ "content": "L _ { \\hat { s t } } [ f ] \\star _ { H } \\psi _ { \\sigma } = L _ { \\hat { s } } L _ { \\hat { t } } [ f ] \\star _ { H } \\psi _ { \\sigma } = L _ { \\hat { s } } [ L _ { \\hat { t } } [ f ] \\star _ { H } \\psi _ { \\sigma } ] = L _ { \\hat { s } } L _ { \\hat { t } } [ f \\star _ { H } \\psi _ { \\sigma } ] = L _ { \\hat { s } \\hat { t } } [ f \\star _ { H } \\psi _ { \\sigma } ] = L _ { \\hat { s t } } [ f \\star _ { H } \\psi _ { \\sigma } ]",
+ "type": "interline_equation",
+ "image_path": "209a5ce9ce9f2fb60134f9570f24b4b4f44e3552f47a5c97a83498b4435f0da6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 119,
+ 402,
+ 475,
+ 416
+ ],
+ "spans": [],
+ "index": 13
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 81,
+ 237,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 239,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 239,
+ 95
+ ],
+ "score": 1.0,
+ "content": "B TIME PERFORMANCE",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 106,
+ 505,
+ 140
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 106,
+ 505,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 106,
+ 505,
+ 118
+ ],
+ "score": 1.0,
+ "content": "We report the average time per epoch of different methods for scale equivariance and local scale",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 117,
+ 505,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 117,
+ 505,
+ 129
+ ],
+ "score": 1.0,
+ "content": "invariance in Table 4. Experimental setups from Section 6.2 are used. We used 1 Nvidia GeForce",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 127,
+ 279,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 127,
+ 279,
+ 140
+ ],
+ "score": 1.0,
+ "content": "GTX 1080Ti GPU for training the models.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 145,
+ 505,
+ 222
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 145,
+ 505,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 145,
+ 505,
+ 159
+ ],
+ "score": 1.0,
+ "content": "The methods relying on image rescaling techniques during training (SiCNN, SI-ConvNet, SEVF)",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 156,
+ 506,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 156,
+ 506,
+ 169
+ ],
+ "score": 1.0,
+ "content": "demonstrate significantly worse time performance that the ones, using either steerable filters or filter",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 166,
+ 505,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 166,
+ 505,
+ 180
+ ],
+ "score": 1.0,
+ "content": "dilation. Additionally, we see that our method outperforms SS-CNN by a wide margin. Despite",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 178,
+ 505,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 178,
+ 505,
+ 190
+ ],
+ "score": 1.0,
+ "content": "the similar filter sizes and comparable number of parameters between SS-CNN and SESN Scalar,",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 190,
+ 505,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 190,
+ 505,
+ 200
+ ],
+ "score": 1.0,
+ "content": "the second one demonstrates significantly better results due to the algorithm proposed in Section 4.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 199,
+ 505,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 199,
+ 505,
+ 213
+ ],
+ "score": 1.0,
+ "content": "Finally, DSS performs slightly faster in some cases than our method as each convolution involves",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 211,
+ 390,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 390,
+ 223
+ ],
+ "score": 1.0,
+ "content": "less FLOPs. Dilated filters are sparse, while steerable filters are dense.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 216,
+ 232,
+ 394,
+ 382
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 216,
+ 232,
+ 394,
+ 382
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 216,
+ 232,
+ 394,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 216,
+ 232,
+ 394,
+ 382
+ ],
+ "score": 0.979,
+ "html": "| Method | 28×28,s | 56 × 56,s |
| CNN | 3.8 | 3.8 |
| SiCNN Scalar | 13.5 | 18.9 |
| SiCNN Vector | 15.3 | 22.8 |
| SI-ConvNet | 18.4 | 33.1 |
| SEVF Scalar | 21.0 | 38.4 |
| SEVF Vector | 25.4 | 46.0 |
| DSS Scalar | 3.9 | 5.0 |
| DSS Vector | 3.9 | 4.8 |
| SS-CNN | 14.8 | 16.6 |
| SESN Scalar | 3.8 | 5.1 |
| SESN Vector | 3.8 | 6.8 |
",
+ "type": "table",
+ "image_path": "93965c7a801a3b764fa0fd4c1064ce418f0949d19c220c38a8d01f4dd5578295.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 15.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 216,
+ 232,
+ 394,
+ 247.0
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 216,
+ 247.0,
+ 394,
+ 262.0
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 216,
+ 262.0,
+ 394,
+ 277.0
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 216,
+ 277.0,
+ 394,
+ 292.0
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 216,
+ 292.0,
+ 394,
+ 307.0
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 216,
+ 307.0,
+ 394,
+ 322.0
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 216,
+ 322.0,
+ 394,
+ 337.0
+ ],
+ "spans": [],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 216,
+ 337.0,
+ 394,
+ 352.0
+ ],
+ "spans": [],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 216,
+ 352.0,
+ 394,
+ 367.0
+ ],
+ "spans": [],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 216,
+ 367.0,
+ 394,
+ 382.0
+ ],
+ "spans": [],
+ "index": 20
+ }
+ ]
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 394,
+ 504,
+ 407
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 393,
+ 505,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 393,
+ 417,
+ 409
+ ],
+ "score": 1.0,
+ "content": "Table 4: Average time per epoch during training on input data with resolution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 417,
+ 395,
+ 451,
+ 406
+ ],
+ "score": 0.91,
+ "content": "2 8 \\times 2 8",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 451,
+ 393,
+ 468,
+ 409
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 468,
+ 395,
+ 501,
+ 406
+ ],
+ "score": 0.88,
+ "content": "5 6 \\times 5 6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 502,
+ 393,
+ 505,
+ 409
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 429,
+ 160,
+ 442
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 162,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 162,
+ 446
+ ],
+ "score": 1.0,
+ "content": "C BASIS",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 454,
+ 506,
+ 477
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 453,
+ 506,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 453,
+ 293,
+ 468
+ ],
+ "score": 1.0,
+ "content": "Assuming that the center of the filter is point",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 293,
+ 455,
+ 316,
+ 467
+ ],
+ "score": 0.92,
+ "content": "( 0 , 0 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 316,
+ 453,
+ 378,
+ 468
+ ],
+ "score": 1.0,
+ "content": "in coordinates",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 379,
+ 455,
+ 402,
+ 467
+ ],
+ "score": 0.92,
+ "content": "( x , y )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 402,
+ 453,
+ 506,
+ 468
+ ],
+ "score": 1.0,
+ "content": ", we use the filters of the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 464,
+ 172,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 464,
+ 172,
+ 479
+ ],
+ "score": 1.0,
+ "content": "following form:",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 197,
+ 475,
+ 413,
+ 501
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 197,
+ 475,
+ 413,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 197,
+ 475,
+ 413,
+ 501
+ ],
+ "score": 0.93,
+ "content": "\\psi _ { \\sigma } ( x , y ) = A \\frac { 1 } { \\sigma ^ { 2 } } H _ { n } \\Big ( \\frac { x } { \\sigma } \\Big ) H _ { m } \\Big ( \\frac { y } { \\sigma } \\Big ) \\exp \\Big [ - \\frac { x ^ { 2 } + y ^ { 2 } } { 2 \\sigma ^ { 2 } } \\Big ]",
+ "type": "interline_equation",
+ "image_path": "12e4e5c2514c596452b622d6ef91e349c9a2b5187c69a2fd93298c8308b935de.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 25,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 197,
+ 475,
+ 413,
+ 501
+ ],
+ "spans": [],
+ "index": 25
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 508,
+ 504,
+ 531
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 507,
+ 505,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 507,
+ 128,
+ 521
+ ],
+ "score": 1.0,
+ "content": "Here",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 128,
+ 509,
+ 137,
+ 518
+ ],
+ "score": 0.8,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 137,
+ 507,
+ 251,
+ 521
+ ],
+ "score": 1.0,
+ "content": "is a constant independent on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 252,
+ 510,
+ 259,
+ 519
+ ],
+ "score": 0.63,
+ "content": "\\sigma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 259,
+ 507,
+ 262,
+ 521
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 263,
+ 509,
+ 277,
+ 519
+ ],
+ "score": 0.8,
+ "content": "H _ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 278,
+ 507,
+ 397,
+ 521
+ ],
+ "score": 1.0,
+ "content": "— Hermite polynomial of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 398,
+ 510,
+ 405,
+ 518
+ ],
+ "score": 0.81,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 405,
+ 507,
+ 505,
+ 521
+ ],
+ "score": 1.0,
+ "content": "-th order. We iterate over",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 519,
+ 386,
+ 531
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 519,
+ 182,
+ 531
+ ],
+ "score": 1.0,
+ "content": "increasing pairs of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 183,
+ 521,
+ 203,
+ 531
+ ],
+ "score": 0.87,
+ "content": "n , m",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 519,
+ 386,
+ 531
+ ],
+ "score": 1.0,
+ "content": "to generate the required number of functions.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5
+ }
+ ],
+ "page_idx": 12,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 26,
+ 294,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 764
+ ],
+ "score": 1.0,
+ "content": "13",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 81,
+ 237,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 239,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 239,
+ 95
+ ],
+ "score": 1.0,
+ "content": "B TIME PERFORMANCE",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 106,
+ 505,
+ 140
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 106,
+ 505,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 106,
+ 505,
+ 118
+ ],
+ "score": 1.0,
+ "content": "We report the average time per epoch of different methods for scale equivariance and local scale",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 117,
+ 505,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 117,
+ 505,
+ 129
+ ],
+ "score": 1.0,
+ "content": "invariance in Table 4. Experimental setups from Section 6.2 are used. We used 1 Nvidia GeForce",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 106,
+ 127,
+ 279,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 127,
+ 279,
+ 140
+ ],
+ "score": 1.0,
+ "content": "GTX 1080Ti GPU for training the models.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2,
+ "bbox_fs": [
+ 106,
+ 106,
+ 505,
+ 140
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 145,
+ 505,
+ 222
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 145,
+ 505,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 145,
+ 505,
+ 159
+ ],
+ "score": 1.0,
+ "content": "The methods relying on image rescaling techniques during training (SiCNN, SI-ConvNet, SEVF)",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 156,
+ 506,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 156,
+ 506,
+ 169
+ ],
+ "score": 1.0,
+ "content": "demonstrate significantly worse time performance that the ones, using either steerable filters or filter",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 166,
+ 505,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 166,
+ 505,
+ 180
+ ],
+ "score": 1.0,
+ "content": "dilation. Additionally, we see that our method outperforms SS-CNN by a wide margin. Despite",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 178,
+ 505,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 178,
+ 505,
+ 190
+ ],
+ "score": 1.0,
+ "content": "the similar filter sizes and comparable number of parameters between SS-CNN and SESN Scalar,",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 190,
+ 505,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 190,
+ 505,
+ 200
+ ],
+ "score": 1.0,
+ "content": "the second one demonstrates significantly better results due to the algorithm proposed in Section 4.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 199,
+ 505,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 199,
+ 505,
+ 213
+ ],
+ "score": 1.0,
+ "content": "Finally, DSS performs slightly faster in some cases than our method as each convolution involves",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 211,
+ 390,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 211,
+ 390,
+ 223
+ ],
+ "score": 1.0,
+ "content": "less FLOPs. Dilated filters are sparse, while steerable filters are dense.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 7,
+ "bbox_fs": [
+ 105,
+ 145,
+ 506,
+ 223
+ ]
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 216,
+ 232,
+ 394,
+ 382
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 216,
+ 232,
+ 394,
+ 382
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 216,
+ 232,
+ 394,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 216,
+ 232,
+ 394,
+ 382
+ ],
+ "score": 0.979,
+ "html": "| Method | 28×28,s | 56 × 56,s |
| CNN | 3.8 | 3.8 |
| SiCNN Scalar | 13.5 | 18.9 |
| SiCNN Vector | 15.3 | 22.8 |
| SI-ConvNet | 18.4 | 33.1 |
| SEVF Scalar | 21.0 | 38.4 |
| SEVF Vector | 25.4 | 46.0 |
| DSS Scalar | 3.9 | 5.0 |
| DSS Vector | 3.9 | 4.8 |
| SS-CNN | 14.8 | 16.6 |
| SESN Scalar | 3.8 | 5.1 |
| SESN Vector | 3.8 | 6.8 |
",
+ "type": "table",
+ "image_path": "93965c7a801a3b764fa0fd4c1064ce418f0949d19c220c38a8d01f4dd5578295.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 15.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 216,
+ 232,
+ 394,
+ 247.0
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 216,
+ 247.0,
+ 394,
+ 262.0
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 216,
+ 262.0,
+ 394,
+ 277.0
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 216,
+ 277.0,
+ 394,
+ 292.0
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 216,
+ 292.0,
+ 394,
+ 307.0
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 216,
+ 307.0,
+ 394,
+ 322.0
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 216,
+ 322.0,
+ 394,
+ 337.0
+ ],
+ "spans": [],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 216,
+ 337.0,
+ 394,
+ 352.0
+ ],
+ "spans": [],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 216,
+ 352.0,
+ 394,
+ 367.0
+ ],
+ "spans": [],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 216,
+ 367.0,
+ 394,
+ 382.0
+ ],
+ "spans": [],
+ "index": 20
+ }
+ ]
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 394,
+ 504,
+ 407
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 393,
+ 505,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 393,
+ 417,
+ 409
+ ],
+ "score": 1.0,
+ "content": "Table 4: Average time per epoch during training on input data with resolution",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 417,
+ 395,
+ 451,
+ 406
+ ],
+ "score": 0.91,
+ "content": "2 8 \\times 2 8",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 451,
+ 393,
+ 468,
+ 409
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 468,
+ 395,
+ 501,
+ 406
+ ],
+ "score": 0.88,
+ "content": "5 6 \\times 5 6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 502,
+ 393,
+ 505,
+ 409
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 21,
+ "bbox_fs": [
+ 105,
+ 393,
+ 505,
+ 409
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 429,
+ 160,
+ 442
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 162,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 162,
+ 446
+ ],
+ "score": 1.0,
+ "content": "C BASIS",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 454,
+ 506,
+ 477
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 453,
+ 506,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 453,
+ 293,
+ 468
+ ],
+ "score": 1.0,
+ "content": "Assuming that the center of the filter is point",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 293,
+ 455,
+ 316,
+ 467
+ ],
+ "score": 0.92,
+ "content": "( 0 , 0 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 316,
+ 453,
+ 378,
+ 468
+ ],
+ "score": 1.0,
+ "content": "in coordinates",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 379,
+ 455,
+ 402,
+ 467
+ ],
+ "score": 0.92,
+ "content": "( x , y )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 402,
+ 453,
+ 506,
+ 468
+ ],
+ "score": 1.0,
+ "content": ", we use the filters of the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 464,
+ 172,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 464,
+ 172,
+ 479
+ ],
+ "score": 1.0,
+ "content": "following form:",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23.5,
+ "bbox_fs": [
+ 105,
+ 453,
+ 506,
+ 479
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 197,
+ 475,
+ 413,
+ 501
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 197,
+ 475,
+ 413,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 197,
+ 475,
+ 413,
+ 501
+ ],
+ "score": 0.93,
+ "content": "\\psi _ { \\sigma } ( x , y ) = A \\frac { 1 } { \\sigma ^ { 2 } } H _ { n } \\Big ( \\frac { x } { \\sigma } \\Big ) H _ { m } \\Big ( \\frac { y } { \\sigma } \\Big ) \\exp \\Big [ - \\frac { x ^ { 2 } + y ^ { 2 } } { 2 \\sigma ^ { 2 } } \\Big ]",
+ "type": "interline_equation",
+ "image_path": "12e4e5c2514c596452b622d6ef91e349c9a2b5187c69a2fd93298c8308b935de.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 25,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 197,
+ 475,
+ 413,
+ 501
+ ],
+ "spans": [],
+ "index": 25
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 508,
+ 504,
+ 531
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 507,
+ 505,
+ 521
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 507,
+ 128,
+ 521
+ ],
+ "score": 1.0,
+ "content": "Here",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 128,
+ 509,
+ 137,
+ 518
+ ],
+ "score": 0.8,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 137,
+ 507,
+ 251,
+ 521
+ ],
+ "score": 1.0,
+ "content": "is a constant independent on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 252,
+ 510,
+ 259,
+ 519
+ ],
+ "score": 0.63,
+ "content": "\\sigma",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 259,
+ 507,
+ 262,
+ 521
+ ],
+ "score": 1.0,
+ "content": ",",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 263,
+ 509,
+ 277,
+ 519
+ ],
+ "score": 0.8,
+ "content": "H _ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 278,
+ 507,
+ 397,
+ 521
+ ],
+ "score": 1.0,
+ "content": "— Hermite polynomial of the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 398,
+ 510,
+ 405,
+ 518
+ ],
+ "score": 0.81,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 405,
+ 507,
+ 505,
+ 521
+ ],
+ "score": 1.0,
+ "content": "-th order. We iterate over",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 519,
+ 386,
+ 531
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 519,
+ 182,
+ 531
+ ],
+ "score": 1.0,
+ "content": "increasing pairs of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 183,
+ 521,
+ 203,
+ 531
+ ],
+ "score": 0.87,
+ "content": "n , m",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 519,
+ 386,
+ 531
+ ],
+ "score": 1.0,
+ "content": "to generate the required number of functions.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 105,
+ 507,
+ 505,
+ 531
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 81,
+ 257,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 79,
+ 259,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 79,
+ 259,
+ 96
+ ],
+ "score": 1.0,
+ "content": "D MODEL CONFIGURATION",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 106,
+ 199,
+ 118
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 106,
+ 199,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 106,
+ 199,
+ 119
+ ],
+ "score": 1.0,
+ "content": "D.1 MNIST-SCALE",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 172,
+ 129,
+ 438,
+ 241
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 172,
+ 129,
+ 438,
+ 241
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 172,
+ 129,
+ 438,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 172,
+ 129,
+ 438,
+ 241
+ ],
+ "score": 0.982,
+ "html": "| Method | Conv 1 | Conv 2 | Conv 3 | FC1 | # Scales |
| CNN | 32 | 63 | 95 | | 1 |
| SiCNN | 32 | 63 | 95 | | 7 |
| SI-ConvNet | 32 | 63 | 95 | | 7 |
| SEVF Scalar | 32 | 63 | 95 | 256 | 8 |
| SEVF Vector | 23 | 45 | 68 | | 8 |
| DSS | 32 | 63 | 95 | | 4 |
| SS-CNN | 30 | 60 | 90 | | 6 |
| SESN | 32 | 63 | 95 | | 4 |
",
+ "type": "table",
+ "image_path": "d9d73c12207c1ef8b3a5e527b54d63142e45ebcd045aa8c1607b5bd141358528.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 172,
+ 129,
+ 438,
+ 166.33333333333334
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 172,
+ 166.33333333333334,
+ 438,
+ 203.66666666666669
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 172,
+ 203.66666666666669,
+ 438,
+ 241.00000000000003
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 107,
+ 254,
+ 503,
+ 277
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 267
+ ],
+ "score": 1.0,
+ "content": "Table 5: Number of channels in convolutional layers, number of units in fully-connected layers and",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 264,
+ 337,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 264,
+ 337,
+ 277
+ ],
+ "score": 1.0,
+ "content": "number of scales used by different models in Section 6.2.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5
+ }
+ ],
+ "index": 4.25
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 187,
+ 323,
+ 423,
+ 423
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 299,
+ 169,
+ 311
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 298,
+ 169,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 298,
+ 169,
+ 312
+ ],
+ "score": 1.0,
+ "content": "D.2 STL-10",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 187,
+ 323,
+ 423,
+ 423
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 187,
+ 323,
+ 423,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 187,
+ 323,
+ 423,
+ 423
+ ],
+ "score": 0.977,
+ "html": "| Method | Block 1 | Block 2 | Block 3 | # Scales |
| CNN | 16 | 32 | 64 | 1 |
| SiCNN | 16 | 32 | 64 | 3 |
| SI-ConvNet | 16 | 32 | 64 | 3 |
| SEVF | 11 | 23 | 45 | 3 |
| DSS | 16 | 32 | 64 | 4 |
| SS-CNN | 11 | 22 | 44 | 3 |
| SESN | 16 | 32 | 64 | 3 |
",
+ "type": "table",
+ "image_path": "f8fdcdff123c35d84fbfe62d7502b0295003d46a9c1cd4e515b00e4296be05d0.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 11,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 187,
+ 323,
+ 423,
+ 337.2857142857143
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 187,
+ 337.2857142857143,
+ 423,
+ 351.57142857142856
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 187,
+ 351.57142857142856,
+ 423,
+ 365.85714285714283
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 187,
+ 365.85714285714283,
+ 423,
+ 380.1428571428571
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 187,
+ 380.1428571428571,
+ 423,
+ 394.4285714285714
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 187,
+ 394.4285714285714,
+ 423,
+ 408.71428571428567
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 187,
+ 408.71428571428567,
+ 423,
+ 422.99999999999994
+ ],
+ "spans": [],
+ "index": 14
+ }
+ ]
+ }
+ ],
+ "index": 9.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 436,
+ 505,
+ 459
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 505,
+ 449
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 505,
+ 449
+ ],
+ "score": 1.0,
+ "content": "Table 6: Number of channels in convolutional blocks and number of scales used by different models",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 446,
+ 408,
+ 459
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 408,
+ 459
+ ],
+ "score": 1.0,
+ "content": "in Section 6.2. We report the number of channels up to the widening factor.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15.5
+ }
+ ],
+ "page_idx": 13,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 294,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 294,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Published as a conference paper at ICLR 2020",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 750,
+ 313,
+ 763
+ ],
+ "score": 1.0,
+ "content": "14",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 81,
+ 257,
+ 94
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 79,
+ 259,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 79,
+ 259,
+ 96
+ ],
+ "score": 1.0,
+ "content": "D MODEL CONFIGURATION",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 106,
+ 199,
+ 118
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 106,
+ 199,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 106,
+ 199,
+ 119
+ ],
+ "score": 1.0,
+ "content": "D.1 MNIST-SCALE",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 172,
+ 129,
+ 438,
+ 241
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 172,
+ 129,
+ 438,
+ 241
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 172,
+ 129,
+ 438,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 172,
+ 129,
+ 438,
+ 241
+ ],
+ "score": 0.982,
+ "html": "| Method | Conv 1 | Conv 2 | Conv 3 | FC1 | # Scales |
| CNN | 32 | 63 | 95 | | 1 |
| SiCNN | 32 | 63 | 95 | | 7 |
| SI-ConvNet | 32 | 63 | 95 | | 7 |
| SEVF Scalar | 32 | 63 | 95 | 256 | 8 |
| SEVF Vector | 23 | 45 | 68 | | 8 |
| DSS | 32 | 63 | 95 | | 4 |
| SS-CNN | 30 | 60 | 90 | | 6 |
| SESN | 32 | 63 | 95 | | 4 |
",
+ "type": "table",
+ "image_path": "d9d73c12207c1ef8b3a5e527b54d63142e45ebcd045aa8c1607b5bd141358528.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 172,
+ 129,
+ 438,
+ 166.33333333333334
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 172,
+ 166.33333333333334,
+ 438,
+ 203.66666666666669
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 172,
+ 203.66666666666669,
+ 438,
+ 241.00000000000003
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 107,
+ 254,
+ 503,
+ 277
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 267
+ ],
+ "score": 1.0,
+ "content": "Table 5: Number of channels in convolutional layers, number of units in fully-connected layers and",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 264,
+ 337,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 264,
+ 337,
+ 277
+ ],
+ "score": 1.0,
+ "content": "number of scales used by different models in Section 6.2.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5.5
+ }
+ ],
+ "index": 4.25
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 187,
+ 323,
+ 423,
+ 423
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 299,
+ 169,
+ 311
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 298,
+ 169,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 298,
+ 169,
+ 312
+ ],
+ "score": 1.0,
+ "content": "D.2 STL-10",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 187,
+ 323,
+ 423,
+ 423
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 187,
+ 323,
+ 423,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 187,
+ 323,
+ 423,
+ 423
+ ],
+ "score": 0.977,
+ "html": "| Method | Block 1 | Block 2 | Block 3 | # Scales |
| CNN | 16 | 32 | 64 | 1 |
| SiCNN | 16 | 32 | 64 | 3 |
| SI-ConvNet | 16 | 32 | 64 | 3 |
| SEVF | 11 | 23 | 45 | 3 |
| DSS | 16 | 32 | 64 | 4 |
| SS-CNN | 11 | 22 | 44 | 3 |
| SESN | 16 | 32 | 64 | 3 |
",
+ "type": "table",
+ "image_path": "f8fdcdff123c35d84fbfe62d7502b0295003d46a9c1cd4e515b00e4296be05d0.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 11,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 187,
+ 323,
+ 423,
+ 337.2857142857143
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 187,
+ 337.2857142857143,
+ 423,
+ 351.57142857142856
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 187,
+ 351.57142857142856,
+ 423,
+ 365.85714285714283
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 187,
+ 365.85714285714283,
+ 423,
+ 380.1428571428571
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 187,
+ 380.1428571428571,
+ 423,
+ 394.4285714285714
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 187,
+ 394.4285714285714,
+ 423,
+ 408.71428571428567
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 187,
+ 408.71428571428567,
+ 423,
+ 422.99999999999994
+ ],
+ "spans": [],
+ "index": 14
+ }
+ ]
+ }
+ ],
+ "index": 9.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 436,
+ 505,
+ 459
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 505,
+ 449
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 505,
+ 449
+ ],
+ "score": 1.0,
+ "content": "Table 6: Number of channels in convolutional blocks and number of scales used by different models",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 446,
+ 408,
+ 459
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 446,
+ 408,
+ 459
+ ],
+ "score": 1.0,
+ "content": "in Section 6.2. We report the number of channels up to the widening factor.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15.5,
+ "bbox_fs": [
+ 105,
+ 435,
+ 505,
+ 459
+ ]
+ }
+ ]
+ }
+ ],
+ "_backend": "pipeline",
+ "_version_name": "2.2.2"
+}
\ No newline at end of file