diff --git "a/parse/dev/_LceCyuVcH/_LceCyuVcH_middle.json" "b/parse/dev/_LceCyuVcH/_LceCyuVcH_middle.json" new file mode 100644--- /dev/null +++ "b/parse/dev/_LceCyuVcH/_LceCyuVcH_middle.json" @@ -0,0 +1,37692 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 110, + 97, + 502, + 137 + ], + "lines": [ + { + "bbox": [ + 108, + 96, + 502, + 120 + ], + "spans": [ + { + "bbox": [ + 108, + 96, + 502, + 120 + ], + "score": 1.0, + "content": "Language Models with Image Descriptors are Strong", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 116, + 441, + 140 + ], + "spans": [ + { + "bbox": [ + 169, + 116, + 441, + 140 + ], + "score": 1.0, + "content": "Few-Shot Video-Language Learners", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 154, + 177, + 459, + 237 + ], + "lines": [ + { + "bbox": [ + 167, + 177, + 447, + 192 + ], + "spans": [ + { + "bbox": [ + 167, + 177, + 299, + 192 + ], + "score": 1.0, + "content": "Zhenhailong Wang1∗, Manling", + "type": "text" + }, + { + "bbox": [ + 299, + 178, + 314, + 190 + ], + "score": 0.36, + "content": "\\mathbf { L i } ^ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 177, + 357, + 192 + ], + "score": 1.0, + "content": "∗, Ruochen", + "type": "text" + }, + { + "bbox": [ + 358, + 178, + 376, + 190 + ], + "score": 0.76, + "content": "\\mathbf { X } \\mathbf { u } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 177, + 447, + 192 + ], + "score": 1.0, + "content": ", Luowei Zhou2†,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 190, + 460, + 204 + ], + "spans": [ + { + "bbox": [ + 153, + 190, + 297, + 204 + ], + "score": 1.0, + "content": "Jie Lei3, Xudong Lin4, Shuohang", + "type": "text" + }, + { + "bbox": [ + 298, + 190, + 327, + 202 + ], + "score": 0.31, + "content": "\\mathbf { W a n g } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 190, + 351, + 204 + ], + "score": 1.0, + "content": ", Ziyi", + "type": "text" + }, + { + "bbox": [ + 351, + 191, + 377, + 203 + ], + "score": 0.3, + "content": "\\mathbf { Y a n g } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 190, + 432, + 204 + ], + "score": 1.0, + "content": ", Chenguang", + "type": "text" + }, + { + "bbox": [ + 433, + 190, + 456, + 202 + ], + "score": 0.45, + "content": "\\mathbf { Z } \\mathbf { h } \\mathbf { u } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 190, + 460, + 204 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 182, + 202, + 427, + 215 + ], + "spans": [ + { + "bbox": [ + 182, + 202, + 427, + 215 + ], + "score": 1.0, + "content": "Derek Hoiem1, Shih-Fu Chang4, Mohit Bansal3, Heng Ji1", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 209, + 212, + 401, + 228 + ], + "spans": [ + { + "bbox": [ + 209, + 212, + 401, + 228 + ], + "score": 1.0, + "content": "1UIUC 2MSR 3UNC 4Columbia University", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 231, + 225, + 380, + 238 + ], + "spans": [ + { + "bbox": [ + 231, + 225, + 380, + 238 + ], + "score": 1.0, + "content": "{wangz3,hengji}@illinois.edu", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 283, + 264, + 328, + 278 + ], + "lines": [ + { + "bbox": [ + 281, + 264, + 330, + 279 + ], + "spans": [ + { + "bbox": [ + 281, + 264, + 330, + 279 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 142, + 289, + 469, + 519 + ], + "lines": [ + { + "bbox": [ + 142, + 290, + 469, + 302 + ], + "spans": [ + { + "bbox": [ + 142, + 290, + 469, + 302 + ], + "score": 1.0, + "content": "The goal of this work is to build flexible video-language models that can generalize", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 300, + 469, + 314 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 469, + 314 + ], + "score": 1.0, + "content": "to various video-to-text tasks from few examples. Existing few-shot video-language", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 312, + 470, + 324 + ], + "spans": [ + { + "bbox": [ + 141, + 312, + 470, + 324 + ], + "score": 1.0, + "content": "learners focus exclusively on the encoder, resulting in the absence of a video-to-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 323, + 470, + 335 + ], + "spans": [ + { + "bbox": [ + 141, + 323, + 470, + 335 + ], + "score": 1.0, + "content": "text decoder to handle generative tasks. Video captioners have been pretrained", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 334, + 470, + 346 + ], + "spans": [ + { + "bbox": [ + 141, + 334, + 470, + 346 + ], + "score": 1.0, + "content": "on large-scale video-language datasets, but they rely heavily on finetuning and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 344, + 470, + 357 + ], + "spans": [ + { + "bbox": [ + 141, + 344, + 470, + 357 + ], + "score": 1.0, + "content": "lack the ability to generate text for unseen tasks in a few-shot setting. We propose", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 355, + 470, + 368 + ], + "spans": [ + { + "bbox": [ + 141, + 355, + 470, + 368 + ], + "score": 1.0, + "content": "VidIL, a few-shot Video-language Learner via Image and Language models,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 366, + 470, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 470, + 379 + ], + "score": 1.0, + "content": "which demonstrates strong performance on few-shot video-to-text tasks without the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 376, + 470, + 390 + ], + "spans": [ + { + "bbox": [ + 141, + 376, + 470, + 390 + ], + "score": 1.0, + "content": "necessity of pretraining or finetuning on any video datasets. We use image-language", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 388, + 470, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 388, + 470, + 401 + ], + "score": 1.0, + "content": "models to translate the video content into frame captions, object, attribute, and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 399, + 470, + 411 + ], + "spans": [ + { + "bbox": [ + 141, + 399, + 470, + 411 + ], + "score": 1.0, + "content": "event phrases, and compose them into a temporal-aware template. We then instruct", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 410, + 469, + 423 + ], + "spans": [ + { + "bbox": [ + 141, + 410, + 469, + 423 + ], + "score": 1.0, + "content": "a language model, with a prompt containing a few in-context examples, to generate", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 140, + 421, + 469, + 434 + ], + "spans": [ + { + "bbox": [ + 140, + 421, + 469, + 434 + ], + "score": 1.0, + "content": "a target output from the composed content. The flexibility of prompting allows", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 432, + 469, + 444 + ], + "spans": [ + { + "bbox": [ + 141, + 432, + 469, + 444 + ], + "score": 1.0, + "content": "the model to capture any form of text input, such as automatic speech recognition", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 443, + 469, + 455 + ], + "spans": [ + { + "bbox": [ + 142, + 443, + 469, + 455 + ], + "score": 1.0, + "content": "(ASR) transcripts. Our experiments demonstrate the power of language models in", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 453, + 469, + 465 + ], + "spans": [ + { + "bbox": [ + 141, + 453, + 469, + 465 + ], + "score": 1.0, + "content": "understanding videos on a wide variety of video-language tasks, including video", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 142, + 465, + 470, + 477 + ], + "spans": [ + { + "bbox": [ + 142, + 465, + 470, + 477 + ], + "score": 1.0, + "content": "captioning, video question answering, video caption retrieval, and video future", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 475, + 469, + 487 + ], + "spans": [ + { + "bbox": [ + 141, + 475, + 469, + 487 + ], + "score": 1.0, + "content": "event prediction. Especially, on video future event prediction, our few-shot model", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 142, + 487, + 469, + 498 + ], + "spans": [ + { + "bbox": [ + 142, + 487, + 469, + 498 + ], + "score": 1.0, + "content": "significantly outperforms state-of-the-art supervised models trained on large-scale", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 142, + 497, + 469, + 510 + ], + "spans": [ + { + "bbox": [ + 142, + 497, + 469, + 510 + ], + "score": 1.0, + "content": "video datasets. Code and processed data are publicly available for research purposes", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 509, + 350, + 520 + ], + "spans": [ + { + "bbox": [ + 141, + 509, + 350, + 520 + ], + "score": 1.0, + "content": "at https://github.com/MikeWangWZHL/VidIL.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 107, + 539, + 190, + 553 + ], + "lines": [ + { + "bbox": [ + 105, + 538, + 192, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 192, + 555 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 564, + 505, + 685 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "score": 1.0, + "content": "One major gap between artificial intelligence and human intelligence lies in their abilities to generalize", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 575, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 505, + 587 + ], + "score": 1.0, + "content": "and perform well on new tasks with limited annotations. Recent advances in large-scale pre-trained", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "score": 1.0, + "content": "generative language models [45, 6, 71, 24] have shown promising few-shot capabilities [72, 43, 63] in", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 596, + 507, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 507, + 610 + ], + "score": 1.0, + "content": "understanding natural language. However, few-shot video-language understanding is still in its infancy.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 607, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 621 + ], + "score": 1.0, + "content": "A particular limitation of most recent video-language frameworks [28, 21, 61, 68, 67, 25, 64, 34] is", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 619, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 506, + 631 + ], + "score": 1.0, + "content": "that they are encoder-only, which means they do not have the ability to generate text from videos for", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 629, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 506, + 642 + ], + "score": 1.0, + "content": "purposes such as captioning [62, 57], question answering [60], and future prediction [23]. Meanwhile,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 641, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 505, + 653 + ], + "score": 1.0, + "content": "unified video-language models [36, 49] that are capable of language decoding still rely heavily", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 651, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 505, + 664 + ], + "score": 1.0, + "content": "on finetuning using a large number of manually annotated video-text pairs, therefore cannot adapt", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 662, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 506, + 676 + ], + "score": 1.0, + "content": "quickly to unseen tasks. Few-shot video-to-text decoding is challenging because the natural language", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 673, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 505, + 686 + ], + "score": 1.0, + "content": "supervision for learning video-language representation is typically based on subtitles and automatic", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 35 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 692, + 220, + 713 + ], + "lines": [ + { + "bbox": [ + 118, + 690, + 194, + 703 + ], + "spans": [ + { + "bbox": [ + 118, + 690, + 194, + 703 + ], + "score": 1.0, + "content": "∗Equal contribution.", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 701, + 221, + 715 + ], + "spans": [ + { + "bbox": [ + 118, + 701, + 221, + 715 + ], + "score": 1.0, + "content": "†Currently at Google Brain.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 731, + 385, + 742 + ], + "lines": [ + { + "bbox": [ + 106, + 730, + 385, + 743 + ], + "spans": [ + { + "bbox": [ + 106, + 730, + 385, + 743 + ], + "score": 1.0, + "content": "36th Conference on Neural Information Processing Systems (NeurIPS 2022).", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 110, + 97, + 502, + 137 + ], + "lines": [ + { + "bbox": [ + 108, + 96, + 502, + 120 + ], + "spans": [ + { + "bbox": [ + 108, + 96, + 502, + 120 + ], + "score": 1.0, + "content": "Language Models with Image Descriptors are Strong", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 116, + 441, + 140 + ], + "spans": [ + { + "bbox": [ + 169, + 116, + 441, + 140 + ], + "score": 1.0, + "content": "Few-Shot Video-Language Learners", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 154, + 177, + 459, + 237 + ], + "lines": [ + { + "bbox": [ + 167, + 177, + 447, + 192 + ], + "spans": [ + { + "bbox": [ + 167, + 177, + 299, + 192 + ], + "score": 1.0, + "content": "Zhenhailong Wang1∗, Manling", + "type": "text" + }, + { + "bbox": [ + 299, + 178, + 314, + 190 + ], + "score": 0.36, + "content": "\\mathbf { L i } ^ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 177, + 357, + 192 + ], + "score": 1.0, + "content": "∗, Ruochen", + "type": "text" + }, + { + "bbox": [ + 358, + 178, + 376, + 190 + ], + "score": 0.76, + "content": "\\mathbf { X } \\mathbf { u } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 177, + 447, + 192 + ], + "score": 1.0, + "content": ", Luowei Zhou2†,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 190, + 460, + 204 + ], + "spans": [ + { + "bbox": [ + 153, + 190, + 297, + 204 + ], + "score": 1.0, + "content": "Jie Lei3, Xudong Lin4, Shuohang", + "type": "text" + }, + { + "bbox": [ + 298, + 190, + 327, + 202 + ], + "score": 0.31, + "content": "\\mathbf { W a n g } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 190, + 351, + 204 + ], + "score": 1.0, + "content": ", Ziyi", + "type": "text" + }, + { + "bbox": [ + 351, + 191, + 377, + 203 + ], + "score": 0.3, + "content": "\\mathbf { Y a n g } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 190, + 432, + 204 + ], + "score": 1.0, + "content": ", Chenguang", + "type": "text" + }, + { + "bbox": [ + 433, + 190, + 456, + 202 + ], + "score": 0.45, + "content": "\\mathbf { Z } \\mathbf { h } \\mathbf { u } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 190, + 460, + 204 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 182, + 202, + 427, + 215 + ], + "spans": [ + { + "bbox": [ + 182, + 202, + 427, + 215 + ], + "score": 1.0, + "content": "Derek Hoiem1, Shih-Fu Chang4, Mohit Bansal3, Heng Ji1", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 209, + 212, + 401, + 228 + ], + "spans": [ + { + "bbox": [ + 209, + 212, + 401, + 228 + ], + "score": 1.0, + "content": "1UIUC 2MSR 3UNC 4Columbia University", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 231, + 225, + 380, + 238 + ], + "spans": [ + { + "bbox": [ + 231, + 225, + 380, + 238 + ], + "score": 1.0, + "content": "{wangz3,hengji}@illinois.edu", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4, + "bbox_fs": [ + 153, + 177, + 460, + 238 + ] + }, + { + "type": "title", + "bbox": [ + 283, + 264, + 328, + 278 + ], + "lines": [ + { + "bbox": [ + 281, + 264, + 330, + 279 + ], + "spans": [ + { + "bbox": [ + 281, + 264, + 330, + 279 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 142, + 289, + 469, + 519 + ], + "lines": [ + { + "bbox": [ + 142, + 290, + 469, + 302 + ], + "spans": [ + { + "bbox": [ + 142, + 290, + 469, + 302 + ], + "score": 1.0, + "content": "The goal of this work is to build flexible video-language models that can generalize", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 300, + 469, + 314 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 469, + 314 + ], + "score": 1.0, + "content": "to various video-to-text tasks from few examples. Existing few-shot video-language", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 312, + 470, + 324 + ], + "spans": [ + { + "bbox": [ + 141, + 312, + 470, + 324 + ], + "score": 1.0, + "content": "learners focus exclusively on the encoder, resulting in the absence of a video-to-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 323, + 470, + 335 + ], + "spans": [ + { + "bbox": [ + 141, + 323, + 470, + 335 + ], + "score": 1.0, + "content": "text decoder to handle generative tasks. Video captioners have been pretrained", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 334, + 470, + 346 + ], + "spans": [ + { + "bbox": [ + 141, + 334, + 470, + 346 + ], + "score": 1.0, + "content": "on large-scale video-language datasets, but they rely heavily on finetuning and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 344, + 470, + 357 + ], + "spans": [ + { + "bbox": [ + 141, + 344, + 470, + 357 + ], + "score": 1.0, + "content": "lack the ability to generate text for unseen tasks in a few-shot setting. We propose", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 355, + 470, + 368 + ], + "spans": [ + { + "bbox": [ + 141, + 355, + 470, + 368 + ], + "score": 1.0, + "content": "VidIL, a few-shot Video-language Learner via Image and Language models,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 366, + 470, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 470, + 379 + ], + "score": 1.0, + "content": "which demonstrates strong performance on few-shot video-to-text tasks without the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 376, + 470, + 390 + ], + "spans": [ + { + "bbox": [ + 141, + 376, + 470, + 390 + ], + "score": 1.0, + "content": "necessity of pretraining or finetuning on any video datasets. We use image-language", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 388, + 470, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 388, + 470, + 401 + ], + "score": 1.0, + "content": "models to translate the video content into frame captions, object, attribute, and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 399, + 470, + 411 + ], + "spans": [ + { + "bbox": [ + 141, + 399, + 470, + 411 + ], + "score": 1.0, + "content": "event phrases, and compose them into a temporal-aware template. We then instruct", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 410, + 469, + 423 + ], + "spans": [ + { + "bbox": [ + 141, + 410, + 469, + 423 + ], + "score": 1.0, + "content": "a language model, with a prompt containing a few in-context examples, to generate", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 140, + 421, + 469, + 434 + ], + "spans": [ + { + "bbox": [ + 140, + 421, + 469, + 434 + ], + "score": 1.0, + "content": "a target output from the composed content. The flexibility of prompting allows", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 432, + 469, + 444 + ], + "spans": [ + { + "bbox": [ + 141, + 432, + 469, + 444 + ], + "score": 1.0, + "content": "the model to capture any form of text input, such as automatic speech recognition", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 443, + 469, + 455 + ], + "spans": [ + { + "bbox": [ + 142, + 443, + 469, + 455 + ], + "score": 1.0, + "content": "(ASR) transcripts. Our experiments demonstrate the power of language models in", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 453, + 469, + 465 + ], + "spans": [ + { + "bbox": [ + 141, + 453, + 469, + 465 + ], + "score": 1.0, + "content": "understanding videos on a wide variety of video-language tasks, including video", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 142, + 465, + 470, + 477 + ], + "spans": [ + { + "bbox": [ + 142, + 465, + 470, + 477 + ], + "score": 1.0, + "content": "captioning, video question answering, video caption retrieval, and video future", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 475, + 469, + 487 + ], + "spans": [ + { + "bbox": [ + 141, + 475, + 469, + 487 + ], + "score": 1.0, + "content": "event prediction. Especially, on video future event prediction, our few-shot model", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 142, + 487, + 469, + 498 + ], + "spans": [ + { + "bbox": [ + 142, + 487, + 469, + 498 + ], + "score": 1.0, + "content": "significantly outperforms state-of-the-art supervised models trained on large-scale", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 142, + 497, + 469, + 510 + ], + "spans": [ + { + "bbox": [ + 142, + 497, + 469, + 510 + ], + "score": 1.0, + "content": "video datasets. Code and processed data are publicly available for research purposes", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 509, + 350, + 520 + ], + "spans": [ + { + "bbox": [ + 141, + 509, + 350, + 520 + ], + "score": 1.0, + "content": "at https://github.com/MikeWangWZHL/VidIL.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 18, + "bbox_fs": [ + 140, + 290, + 470, + 520 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 539, + 190, + 553 + ], + "lines": [ + { + "bbox": [ + 105, + 538, + 192, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 192, + 555 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 564, + 505, + 685 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "score": 1.0, + "content": "One major gap between artificial intelligence and human intelligence lies in their abilities to generalize", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 575, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 505, + 587 + ], + "score": 1.0, + "content": "and perform well on new tasks with limited annotations. Recent advances in large-scale pre-trained", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "score": 1.0, + "content": "generative language models [45, 6, 71, 24] have shown promising few-shot capabilities [72, 43, 63] in", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 596, + 507, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 507, + 610 + ], + "score": 1.0, + "content": "understanding natural language. However, few-shot video-language understanding is still in its infancy.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 607, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 621 + ], + "score": 1.0, + "content": "A particular limitation of most recent video-language frameworks [28, 21, 61, 68, 67, 25, 64, 34] is", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 619, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 506, + 631 + ], + "score": 1.0, + "content": "that they are encoder-only, which means they do not have the ability to generate text from videos for", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 629, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 506, + 642 + ], + "score": 1.0, + "content": "purposes such as captioning [62, 57], question answering [60], and future prediction [23]. Meanwhile,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 641, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 505, + 653 + ], + "score": 1.0, + "content": "unified video-language models [36, 49] that are capable of language decoding still rely heavily", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 651, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 505, + 664 + ], + "score": 1.0, + "content": "on finetuning using a large number of manually annotated video-text pairs, therefore cannot adapt", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 662, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 506, + 676 + ], + "score": 1.0, + "content": "quickly to unseen tasks. Few-shot video-to-text decoding is challenging because the natural language", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 673, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 505, + 686 + ], + "score": 1.0, + "content": "supervision for learning video-language representation is typically based on subtitles and automatic", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "speech recognition (ASR) transcripts [39, 68], which differ significantly from downstream tasks in", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 84, + 483, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 84, + 483, + 96 + ], + "score": 1.0, + "content": "terms of distribution and may have poor semantic alignment across vision and text modalities.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 563, + 507, + 686 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 504, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "speech recognition (ASR) transcripts [39, 68], which differ significantly from downstream tasks in", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 84, + 483, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 84, + 483, + 96 + ], + "score": 1.0, + "content": "terms of distribution and may have poor semantic alignment across vision and text modalities.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 505, + 177 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 505, + 114 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 114 + ], + "score": 1.0, + "content": "We propose to address this problem by harnessing the few-shot power of frozen large-scale language", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "models, such as InstructGPT [40]. Our inspiration is derived from the fact that humans are excellent", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 122, + 506, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 506, + 134 + ], + "score": 1.0, + "content": "visual storytellers [15], with the ability to piece together a coherent story from a few isolated images.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 131, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 505, + 146 + ], + "score": 1.0, + "content": "To mimic this, we propose VidIL, a few-shot Video-language Learner via Image and Language", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 144, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 155 + ], + "score": 1.0, + "content": "models, to use image models to provide information about the visual content in the video (as well", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 155, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 506, + 167 + ], + "score": 1.0, + "content": "as optionally use ASR to represent speech), and then we instruct language models to generate a", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 165, + 452, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 452, + 179 + ], + "score": 1.0, + "content": "video-based summary, answer, or other target output for diverse video-language tasks.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 183, + 297, + 389 + ], + "lines": [ + { + "bbox": [ + 106, + 181, + 297, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 297, + 194 + ], + "score": 1.0, + "content": "The main challenge of understanding videos", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 193, + 298, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 298, + 204 + ], + "score": 1.0, + "content": "is that, videos contain rich semantics and tem-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 204, + 297, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 297, + 215 + ], + "score": 1.0, + "content": "poral content at multiple granularities. Unlike", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 214, + 297, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 297, + 227 + ], + "score": 1.0, + "content": "static images which depict objects, attributes", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 225, + 298, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 298, + 237 + ], + "score": 1.0, + "content": "and events in a snapshot, the temporal dimen-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 236, + 297, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 297, + 248 + ], + "score": 1.0, + "content": "sion of videos further conveys the state changes", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 248, + 298, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 298, + 259 + ], + "score": 1.0, + "content": "of the objects, actions, and events. For exam-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 258, + 298, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 258, + 298, + 270 + ], + "score": 1.0, + "content": "ple, in Figure 1, the individual frame captions", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 269, + 299, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 299, + 281 + ], + "score": 1.0, + "content": "of the video clip only describe static visual fea-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 280, + 298, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 298, + 293 + ], + "score": 1.0, + "content": "tures such as \"a person holding a green object", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 291, + 299, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 299, + 303 + ], + "score": 1.0, + "content": "in hand\". In contrast, a correct video-level de-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 302, + 298, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 298, + 313 + ], + "score": 1.0, + "content": "scription would be \"a woman makes realistic", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 313, + 297, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 297, + 324 + ], + "score": 1.0, + "content": "looking leaves and flowers for a cake\", which", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 324, + 297, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 297, + 336 + ], + "score": 1.0, + "content": "involves reasoning over a collection of objects", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 335, + 297, + 346 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 297, + 346 + ], + "score": 1.0, + "content": "and events that occur at different timestamps in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 345, + 298, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 298, + 357 + ], + "score": 1.0, + "content": "the video clip, such as \"cake decorating\" and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 356, + 298, + 369 + ], + "spans": [ + { + "bbox": [ + 107, + 356, + 298, + 369 + ], + "score": 1.0, + "content": "\"flowered design\". Hence, to inform video-level", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 368, + 297, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 297, + 379 + ], + "score": 1.0, + "content": "description and queries, we need to represent all", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 377, + 289, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 289, + 391 + ], + "score": 1.0, + "content": "of this information and its temporal ordering.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 311, + 200, + 497, + 355 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 311, + 200, + 497, + 355 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 311, + 200, + 497, + 355 + ], + "spans": [ + { + "bbox": [ + 311, + 200, + 497, + 355 + ], + "score": 0.947, + "type": "image", + "image_path": "b5f930daf9032229a0d0a4e399381948a20cacaccc02903750ded88494988f0e.jpg" + } + ] + } + ], + "index": 33.5, + "virtual_lines": [ + { + "bbox": [ + 311, + 200, + 497, + 212.91666666666666 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 311, + 212.91666666666666, + 497, + 225.83333333333331 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 311, + 225.83333333333331, + 497, + 238.74999999999997 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 311, + 238.74999999999997, + 497, + 251.66666666666663 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 311, + 251.66666666666663, + 497, + 264.5833333333333 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 311, + 264.5833333333333, + 497, + 277.5 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 311, + 277.5, + 497, + 290.4166666666667 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 311, + 290.4166666666667, + 497, + 303.33333333333337 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 311, + 303.33333333333337, + 497, + 316.25000000000006 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 311, + 316.25000000000006, + 497, + 329.16666666666674 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 311, + 329.16666666666674, + 497, + 342.0833333333334 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 311, + 342.0833333333334, + 497, + 355.0000000000001 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 306, + 367, + 503, + 379 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 303, + 366, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 303, + 366, + 505, + 380 + ], + "score": 1.0, + "content": "Figure 1: Multiple levels of information in videos.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + } + ], + "index": 36.75 + }, + { + "type": "text", + "bbox": [ + 107, + 395, + 505, + 558 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 407 + ], + "score": 1.0, + "content": "To address the unique challenges of videos, we propose to decompose a video into three levels: the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 405, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 506, + 418 + ], + "score": 1.0, + "content": "video output, frame captions, and visual tokens (including objects, events, attributes). One major", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "benefit from this hierarchical video representation is that we can separate the visual and temporal", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 427, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 506, + 441 + ], + "score": 1.0, + "content": "dimensions of a video. We leverage frozen image-language foundational models at lower levels to", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 437, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 506, + 451 + ], + "score": 1.0, + "content": "collect salient visual features from the sparsely sampled frames. Specifically, we first leverage a", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 449, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 461 + ], + "score": 1.0, + "content": "pretrained image-language contrastive model CLIP [44] to perform visual tokenization, based on", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "score": 1.0, + "content": "the similarity score between frames and tokens of objects, events and attributes. The tokenization is", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 471, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 505, + 483 + ], + "score": 1.0, + "content": "done under the guidance of semantics role labeling [14], which provides us with candidate events", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 482, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 494 + ], + "score": 1.0, + "content": "with involved objects and related attributes. Next, in order to capture the overall semantics at the", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 493, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 505 + ], + "score": 1.0, + "content": "frame level, we employ the pretrained image captioner in the image-language model BLIP [26]", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 505, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 516 + ], + "score": 1.0, + "content": "to obtain frame captions. Finally, we instruct a pretrained large language model using in-context", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "score": 1.0, + "content": "learning [40, 13, 51, 48] to interpret visual tokens and frame captions into the target textual output.", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "score": 1.0, + "content": "In detail, we temporally order visual tokens and frame captions using specially designed prompts", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "such as “First...Then...Finally”, to instruct the pretrained language model to track the changes of", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 547, + 417, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 417, + 560 + ], + "score": 1.0, + "content": "objects, events, attributes and frame semantics along the temporal dimension.", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 48 + }, + { + "type": "text", + "bbox": [ + 107, + 564, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "score": 1.0, + "content": "Without pretraining or finetuning on any video datasets, we show that our approach outperforms", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 575, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 505, + 587 + ], + "score": 1.0, + "content": "both video-language and image-language state-of-the-art baselines on few-shot video captioning and", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 586, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 598 + ], + "score": 1.0, + "content": "question answering tasks. Moreover, on video-language event prediction, our approach significantly", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "score": 1.0, + "content": "outperforms fully-supervised models while using only 10 labeled examples. We further demonstrate", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "that our generative model can benefit broader video-language understanding tasks, such as text-video", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "retrieval, via pseudo label generation. Additionally, we show that our model is highly flexible in", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 106, + 630, + 301, + 642 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 301, + 642 + ], + "score": 1.0, + "content": "adding new modalities, such as ASR transcripts.", + "type": "text" + } + ], + "index": 62 + } + ], + "index": 59 + }, + { + "type": "title", + "bbox": [ + 107, + 655, + 197, + 668 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 198, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 198, + 670 + ], + "score": 1.0, + "content": "2 Related Work", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 63 + }, + { + "type": "title", + "bbox": [ + 108, + 679, + 449, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 679, + 450, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 450, + 693 + ], + "score": 1.0, + "content": "2.1 Image-Language Models and Their Applications on Video-Language Tasks", + "type": "text" + } + ], + "index": 64 + } + ], + "index": 64 + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "Large-scale image-language pretraining models optimize image-text matching through contrastive", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 105, + 709, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 724 + ], + "score": 1.0, + "content": "learning [44, 17] and multimodal fusion [65, 27, 58, 66, 35, 52, 8, 29, 73, 70, 18, 16]. Recently,", + "type": "text" + } + ], + "index": 66 + } + ], + "index": 65.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 741, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 741, + 310, + 753 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 504, + 95 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 72, + 506, + 96 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 505, + 177 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 505, + 114 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 114 + ], + "score": 1.0, + "content": "We propose to address this problem by harnessing the few-shot power of frozen large-scale language", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "models, such as InstructGPT [40]. Our inspiration is derived from the fact that humans are excellent", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 122, + 506, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 506, + 134 + ], + "score": 1.0, + "content": "visual storytellers [15], with the ability to piece together a coherent story from a few isolated images.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 131, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 505, + 146 + ], + "score": 1.0, + "content": "To mimic this, we propose VidIL, a few-shot Video-language Learner via Image and Language", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 144, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 505, + 155 + ], + "score": 1.0, + "content": "models, to use image models to provide information about the visual content in the video (as well", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 155, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 506, + 167 + ], + "score": 1.0, + "content": "as optionally use ASR to represent speech), and then we instruct language models to generate a", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 165, + 452, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 452, + 179 + ], + "score": 1.0, + "content": "video-based summary, answer, or other target output for diverse video-language tasks.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 99, + 506, + 179 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 183, + 297, + 389 + ], + "lines": [ + { + "bbox": [ + 106, + 181, + 297, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 297, + 194 + ], + "score": 1.0, + "content": "The main challenge of understanding videos", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 193, + 298, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 298, + 204 + ], + "score": 1.0, + "content": "is that, videos contain rich semantics and tem-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 204, + 297, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 297, + 215 + ], + "score": 1.0, + "content": "poral content at multiple granularities. Unlike", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 214, + 297, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 297, + 227 + ], + "score": 1.0, + "content": "static images which depict objects, attributes", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 225, + 298, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 298, + 237 + ], + "score": 1.0, + "content": "and events in a snapshot, the temporal dimen-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 236, + 297, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 297, + 248 + ], + "score": 1.0, + "content": "sion of videos further conveys the state changes", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 248, + 298, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 298, + 259 + ], + "score": 1.0, + "content": "of the objects, actions, and events. For exam-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 258, + 298, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 258, + 298, + 270 + ], + "score": 1.0, + "content": "ple, in Figure 1, the individual frame captions", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 269, + 299, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 299, + 281 + ], + "score": 1.0, + "content": "of the video clip only describe static visual fea-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 280, + 298, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 298, + 293 + ], + "score": 1.0, + "content": "tures such as \"a person holding a green object", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 291, + 299, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 299, + 303 + ], + "score": 1.0, + "content": "in hand\". In contrast, a correct video-level de-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 302, + 298, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 298, + 313 + ], + "score": 1.0, + "content": "scription would be \"a woman makes realistic", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 313, + 297, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 297, + 324 + ], + "score": 1.0, + "content": "looking leaves and flowers for a cake\", which", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 324, + 297, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 297, + 336 + ], + "score": 1.0, + "content": "involves reasoning over a collection of objects", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 335, + 297, + 346 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 297, + 346 + ], + "score": 1.0, + "content": "and events that occur at different timestamps in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 345, + 298, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 298, + 357 + ], + "score": 1.0, + "content": "the video clip, such as \"cake decorating\" and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 107, + 356, + 298, + 369 + ], + "spans": [ + { + "bbox": [ + 107, + 356, + 298, + 369 + ], + "score": 1.0, + "content": "\"flowered design\". Hence, to inform video-level", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 368, + 297, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 297, + 379 + ], + "score": 1.0, + "content": "description and queries, we need to represent all", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 377, + 289, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 289, + 391 + ], + "score": 1.0, + "content": "of this information and its temporal ordering.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 181, + 299, + 391 + ] + }, + { + "type": "image", + "bbox": [ + 311, + 200, + 497, + 355 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 311, + 200, + 497, + 355 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 311, + 200, + 497, + 355 + ], + "spans": [ + { + "bbox": [ + 311, + 200, + 497, + 355 + ], + "score": 0.947, + "type": "image", + "image_path": "b5f930daf9032229a0d0a4e399381948a20cacaccc02903750ded88494988f0e.jpg" + } + ] + } + ], + "index": 33.5, + "virtual_lines": [ + { + "bbox": [ + 311, + 200, + 497, + 212.91666666666666 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 311, + 212.91666666666666, + 497, + 225.83333333333331 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 311, + 225.83333333333331, + 497, + 238.74999999999997 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 311, + 238.74999999999997, + 497, + 251.66666666666663 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 311, + 251.66666666666663, + 497, + 264.5833333333333 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 311, + 264.5833333333333, + 497, + 277.5 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 311, + 277.5, + 497, + 290.4166666666667 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 311, + 290.4166666666667, + 497, + 303.33333333333337 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 311, + 303.33333333333337, + 497, + 316.25000000000006 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 311, + 316.25000000000006, + 497, + 329.16666666666674 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 311, + 329.16666666666674, + 497, + 342.0833333333334 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 311, + 342.0833333333334, + 497, + 355.0000000000001 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 306, + 367, + 503, + 379 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 303, + 366, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 303, + 366, + 505, + 380 + ], + "score": 1.0, + "content": "Figure 1: Multiple levels of information in videos.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + } + ], + "index": 36.75 + }, + { + "type": "text", + "bbox": [ + 107, + 395, + 505, + 558 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 407 + ], + "score": 1.0, + "content": "To address the unique challenges of videos, we propose to decompose a video into three levels: the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 405, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 506, + 418 + ], + "score": 1.0, + "content": "video output, frame captions, and visual tokens (including objects, events, attributes). One major", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "benefit from this hierarchical video representation is that we can separate the visual and temporal", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 427, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 506, + 441 + ], + "score": 1.0, + "content": "dimensions of a video. We leverage frozen image-language foundational models at lower levels to", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 437, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 506, + 451 + ], + "score": 1.0, + "content": "collect salient visual features from the sparsely sampled frames. Specifically, we first leverage a", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 449, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 461 + ], + "score": 1.0, + "content": "pretrained image-language contrastive model CLIP [44] to perform visual tokenization, based on", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "score": 1.0, + "content": "the similarity score between frames and tokens of objects, events and attributes. The tokenization is", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 471, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 505, + 483 + ], + "score": 1.0, + "content": "done under the guidance of semantics role labeling [14], which provides us with candidate events", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 482, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 494 + ], + "score": 1.0, + "content": "with involved objects and related attributes. Next, in order to capture the overall semantics at the", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 493, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 505 + ], + "score": 1.0, + "content": "frame level, we employ the pretrained image captioner in the image-language model BLIP [26]", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 505, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 516 + ], + "score": 1.0, + "content": "to obtain frame captions. Finally, we instruct a pretrained large language model using in-context", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "score": 1.0, + "content": "learning [40, 13, 51, 48] to interpret visual tokens and frame captions into the target textual output.", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "score": 1.0, + "content": "In detail, we temporally order visual tokens and frame captions using specially designed prompts", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "such as “First...Then...Finally”, to instruct the pretrained language model to track the changes of", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 547, + 417, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 417, + 560 + ], + "score": 1.0, + "content": "objects, events, attributes and frame semantics along the temporal dimension.", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 48, + "bbox_fs": [ + 105, + 394, + 506, + 560 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 564, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "score": 1.0, + "content": "Without pretraining or finetuning on any video datasets, we show that our approach outperforms", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 575, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 505, + 587 + ], + "score": 1.0, + "content": "both video-language and image-language state-of-the-art baselines on few-shot video captioning and", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 586, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 598 + ], + "score": 1.0, + "content": "question answering tasks. Moreover, on video-language event prediction, our approach significantly", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 609 + ], + "score": 1.0, + "content": "outperforms fully-supervised models while using only 10 labeled examples. We further demonstrate", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "that our generative model can benefit broader video-language understanding tasks, such as text-video", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "retrieval, via pseudo label generation. Additionally, we show that our model is highly flexible in", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 106, + 630, + 301, + 642 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 301, + 642 + ], + "score": 1.0, + "content": "adding new modalities, such as ASR transcripts.", + "type": "text" + } + ], + "index": 62 + } + ], + "index": 59, + "bbox_fs": [ + 105, + 563, + 505, + 642 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 655, + 197, + 668 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 198, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 198, + 670 + ], + "score": 1.0, + "content": "2 Related Work", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 63 + }, + { + "type": "title", + "bbox": [ + 108, + 679, + 449, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 679, + 450, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 450, + 693 + ], + "score": 1.0, + "content": "2.1 Image-Language Models and Their Applications on Video-Language Tasks", + "type": "text" + } + ], + "index": 64 + } + ], + "index": 64 + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "Large-scale image-language pretraining models optimize image-text matching through contrastive", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 105, + 709, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 724 + ], + "score": 1.0, + "content": "learning [44, 17] and multimodal fusion [65, 27, 58, 66, 35, 52, 8, 29, 73, 70, 18, 16]. Recently,", + "type": "text" + } + ], + "index": 66 + } + ], + "index": 65.5, + "bbox_fs": [ + 105, + 699, + 506, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 73, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "BLIP [26] proposes a bootstrapping image-language pretraining framework with a captioner and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 506, + 97 + ], + "score": 1.0, + "content": "a filterer which has shown promising performance on various image-language tasks. However,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 94, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 107 + ], + "score": 1.0, + "content": "video-language pretraining [25, 36, 28, 38, 3, 1, 42, 33] is still hindered by noisy and domain-specific", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 104, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 119 + ], + "score": 1.0, + "content": "video datasets [74, 22, 39]. Naturally, researchers start to explore transferring the rich knowledge", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 116, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 128 + ], + "score": 1.0, + "content": "from image models to videos. Different from the traditional way of representing videos by 3D", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 127, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 506, + 140 + ], + "score": 1.0, + "content": "dense features [12], recent work [21, 25] proves that sparse sampling is an effective way to represent", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 137, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 151 + ], + "score": 1.0, + "content": "videos, which facilitates applying pre-trained image-language models to video-language tasks [37, 11].", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 162 + ], + "score": 1.0, + "content": "Specifically, the image-language model BLIP [26] sets new state-of-the-art on zero-shot retrieval-style", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 160, + 506, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 506, + 172 + ], + "score": 1.0, + "content": "video-language tasks, such as video retrieval and video question answering. However, for generation-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "style tasks such as domain-specific video captioning, video-language model UniVL [36] still leads", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 181, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 506, + 195 + ], + "score": 1.0, + "content": "the performance but highly rely on fine-tuning. In this work, we extend the idea of leveraging", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "image-language models to a wide variety of video-to-text generation tasks. We further connect image-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 203, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 506, + 216 + ], + "score": 1.0, + "content": "language models with language models which empowers our model with strong generalization ability.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 214, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 505, + 229 + ], + "score": 1.0, + "content": "We show that the knowledge from both image-language pretraining and language-only pretraining", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 226, + 351, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 351, + 238 + ], + "score": 1.0, + "content": "can benefit video-language understanding in various aspects.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 7 + }, + { + "type": "title", + "bbox": [ + 107, + 258, + 348, + 271 + ], + "lines": [ + { + "bbox": [ + 104, + 256, + 349, + 274 + ], + "spans": [ + { + "bbox": [ + 104, + 256, + 349, + 274 + ], + "score": 1.0, + "content": "2.2 Unifying MultiModal Tasks with Language Models", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 282, + 505, + 469 + ], + "lines": [ + { + "bbox": [ + 106, + 282, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 506, + 295 + ], + "score": 1.0, + "content": "The community has paid much attention to connecting different modalities with a unified representa-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 292, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 308 + ], + "score": 1.0, + "content": "tion recently. Text-only generation models, such as T5 [46], have been extended to vision-language", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 303, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 318 + ], + "score": 1.0, + "content": "tasks by text generation conditioned on visual features [9, 53, 50, 75, 55]. In order to fully leverage", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 315, + 506, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 506, + 328 + ], + "score": 1.0, + "content": "the generalization power from pretained language models, [63] represents images using text in a", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "fully symbolic way. [32] includes more modalities such as video and audio, but requires annotated", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 337, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 506, + 350 + ], + "score": 1.0, + "content": "video-text data to jointly training the language model with the video and audio tokenizer. In this work,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 348, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 506, + 361 + ], + "score": 1.0, + "content": "we propose a temporal-aware hierarchical representation for describing a video textually. To our", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 358, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 506, + 372 + ], + "score": 1.0, + "content": "knowledge, we are the first work to leverage prompting a frozen language model for tackling few-shot", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "video-language tasks with a unified textual representation. Concurrent work Socratic [69] uses a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "zero-shot language-based world-state history to represent long videos with given time stamps, while", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 391, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 506, + 405 + ], + "score": 1.0, + "content": "our model can quickly adapt to different video and text distributions with few examples. Furthermore,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "we show that by injecting temporal markers to the prompt we can make a pre-trained language model", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 414, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 506, + 426 + ], + "score": 1.0, + "content": "understand fine-grained temporal dynamics in video events. Compared with the concurrent work", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 506, + 437 + ], + "score": 1.0, + "content": "Flamingo [2], which requires dedicated vision-language post-pretraining, our framework does not", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 435, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 506, + 447 + ], + "score": 1.0, + "content": "require to pretrain or finetune on any video data. Our framework is simple and highly modulated", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 445, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 506, + 460 + ], + "score": 1.0, + "content": "where all the components are publicly available. Additionally, our framework is more flexible on", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 455, + 503, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 503, + 471 + ], + "score": 1.0, + "content": "adding new modalities, e.g., automatic speech recognition, without the need for complex redesigning.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 24 + }, + { + "type": "title", + "bbox": [ + 107, + 492, + 165, + 506 + ], + "lines": [ + { + "bbox": [ + 104, + 491, + 167, + 507 + ], + "spans": [ + { + "bbox": [ + 104, + 491, + 167, + 507 + ], + "score": 1.0, + "content": "3 Method", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 505, + 536 + ], + "score": 1.0, + "content": "We propose a hierarchical video representation framework which decomposes a video into three", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "levels, i.e., visual token level, frame level and video level. The motivation is to separate the spatial", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "and temporal dimension of a video in order to leverage image-language and language-only foundation", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "models, such as CLIP [44] and GPT-3 [6]. All three levels use a unified textual representation which", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 567, + 450, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 450, + 579 + ], + "score": 1.0, + "content": "enables us to leverage the powerful few-shot ability from pretrained language models.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + }, + { + "type": "title", + "bbox": [ + 108, + 599, + 267, + 612 + ], + "lines": [ + { + "bbox": [ + 105, + 597, + 269, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 269, + 614 + ], + "score": 1.0, + "content": "3.1 Frame Level: Image Captioning", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 623, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 623, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 505, + 636 + ], + "score": 1.0, + "content": "Following [21] we first perform sparse sampling to obtain several video frames. Unless otherwise", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 634, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 505, + 646 + ], + "score": 1.0, + "content": "specified, we sample 4 frames for frame level and 8 frames for visual token level. We then feed each", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 646, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 505, + 657 + ], + "score": 1.0, + "content": "frame into a pre-trained image-language model to obtain frame level captions. An example can be", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 654, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 670 + ], + "score": 1.0, + "content": "found in the blue part of Figure 2. In our experiments, we use BLIP [26], a recent image-language", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "framework containing both image-grounded encoder and decoder, for generating frame captions. We", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "score": 1.0, + "content": "follow [26] to do both captioning and filtering on each frame. However, as mentioned in Section 1,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "videos contain rich semantics and temporal contents at multiple granularities. It is not enough to", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "generate video-level target text such as video captions solely based on frame captions. Thus, we", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 710, + 470, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 470, + 724 + ], + "score": 1.0, + "content": "further perform visual tokenization for each frame to capture features at a finer granularity.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 44 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 309, + 752 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 73, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "BLIP [26] proposes a bootstrapping image-language pretraining framework with a captioner and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 506, + 97 + ], + "score": 1.0, + "content": "a filterer which has shown promising performance on various image-language tasks. However,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 94, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 107 + ], + "score": 1.0, + "content": "video-language pretraining [25, 36, 28, 38, 3, 1, 42, 33] is still hindered by noisy and domain-specific", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 104, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 119 + ], + "score": 1.0, + "content": "video datasets [74, 22, 39]. Naturally, researchers start to explore transferring the rich knowledge", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 116, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 128 + ], + "score": 1.0, + "content": "from image models to videos. Different from the traditional way of representing videos by 3D", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 127, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 506, + 140 + ], + "score": 1.0, + "content": "dense features [12], recent work [21, 25] proves that sparse sampling is an effective way to represent", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 137, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 151 + ], + "score": 1.0, + "content": "videos, which facilitates applying pre-trained image-language models to video-language tasks [37, 11].", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 162 + ], + "score": 1.0, + "content": "Specifically, the image-language model BLIP [26] sets new state-of-the-art on zero-shot retrieval-style", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 160, + 506, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 506, + 172 + ], + "score": 1.0, + "content": "video-language tasks, such as video retrieval and video question answering. However, for generation-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "style tasks such as domain-specific video captioning, video-language model UniVL [36] still leads", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 181, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 506, + 195 + ], + "score": 1.0, + "content": "the performance but highly rely on fine-tuning. In this work, we extend the idea of leveraging", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "image-language models to a wide variety of video-to-text generation tasks. We further connect image-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 203, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 506, + 216 + ], + "score": 1.0, + "content": "language models with language models which empowers our model with strong generalization ability.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 214, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 505, + 229 + ], + "score": 1.0, + "content": "We show that the knowledge from both image-language pretraining and language-only pretraining", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 226, + 351, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 351, + 238 + ], + "score": 1.0, + "content": "can benefit video-language understanding in various aspects.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 72, + 506, + 238 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 258, + 348, + 271 + ], + "lines": [ + { + "bbox": [ + 104, + 256, + 349, + 274 + ], + "spans": [ + { + "bbox": [ + 104, + 256, + 349, + 274 + ], + "score": 1.0, + "content": "2.2 Unifying MultiModal Tasks with Language Models", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 282, + 505, + 469 + ], + "lines": [ + { + "bbox": [ + 106, + 282, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 506, + 295 + ], + "score": 1.0, + "content": "The community has paid much attention to connecting different modalities with a unified representa-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 292, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 308 + ], + "score": 1.0, + "content": "tion recently. Text-only generation models, such as T5 [46], have been extended to vision-language", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 303, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 318 + ], + "score": 1.0, + "content": "tasks by text generation conditioned on visual features [9, 53, 50, 75, 55]. In order to fully leverage", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 315, + 506, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 506, + 328 + ], + "score": 1.0, + "content": "the generalization power from pretained language models, [63] represents images using text in a", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "fully symbolic way. [32] includes more modalities such as video and audio, but requires annotated", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 337, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 506, + 350 + ], + "score": 1.0, + "content": "video-text data to jointly training the language model with the video and audio tokenizer. In this work,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 348, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 506, + 361 + ], + "score": 1.0, + "content": "we propose a temporal-aware hierarchical representation for describing a video textually. To our", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 358, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 506, + 372 + ], + "score": 1.0, + "content": "knowledge, we are the first work to leverage prompting a frozen language model for tackling few-shot", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "video-language tasks with a unified textual representation. Concurrent work Socratic [69] uses a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "zero-shot language-based world-state history to represent long videos with given time stamps, while", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 391, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 506, + 405 + ], + "score": 1.0, + "content": "our model can quickly adapt to different video and text distributions with few examples. Furthermore,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "we show that by injecting temporal markers to the prompt we can make a pre-trained language model", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 414, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 506, + 426 + ], + "score": 1.0, + "content": "understand fine-grained temporal dynamics in video events. Compared with the concurrent work", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 506, + 437 + ], + "score": 1.0, + "content": "Flamingo [2], which requires dedicated vision-language post-pretraining, our framework does not", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 435, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 506, + 447 + ], + "score": 1.0, + "content": "require to pretrain or finetune on any video data. Our framework is simple and highly modulated", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 445, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 506, + 460 + ], + "score": 1.0, + "content": "where all the components are publicly available. Additionally, our framework is more flexible on", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 455, + 503, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 503, + 471 + ], + "score": 1.0, + "content": "adding new modalities, e.g., automatic speech recognition, without the need for complex redesigning.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 282, + 506, + 471 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 492, + 165, + 506 + ], + "lines": [ + { + "bbox": [ + 104, + 491, + 167, + 507 + ], + "spans": [ + { + "bbox": [ + 104, + 491, + 167, + 507 + ], + "score": 1.0, + "content": "3 Method", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 505, + 536 + ], + "score": 1.0, + "content": "We propose a hierarchical video representation framework which decomposes a video into three", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "levels, i.e., visual token level, frame level and video level. The motivation is to separate the spatial", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "and temporal dimension of a video in order to leverage image-language and language-only foundation", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "models, such as CLIP [44] and GPT-3 [6]. All three levels use a unified textual representation which", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 567, + 450, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 450, + 579 + ], + "score": 1.0, + "content": "enables us to leverage the powerful few-shot ability from pretrained language models.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 522, + 505, + 579 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 599, + 267, + 612 + ], + "lines": [ + { + "bbox": [ + 105, + 597, + 269, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 269, + 614 + ], + "score": 1.0, + "content": "3.1 Frame Level: Image Captioning", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 623, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 623, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 505, + 636 + ], + "score": 1.0, + "content": "Following [21] we first perform sparse sampling to obtain several video frames. Unless otherwise", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 634, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 505, + 646 + ], + "score": 1.0, + "content": "specified, we sample 4 frames for frame level and 8 frames for visual token level. We then feed each", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 646, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 505, + 657 + ], + "score": 1.0, + "content": "frame into a pre-trained image-language model to obtain frame level captions. An example can be", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 654, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 670 + ], + "score": 1.0, + "content": "found in the blue part of Figure 2. In our experiments, we use BLIP [26], a recent image-language", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "framework containing both image-grounded encoder and decoder, for generating frame captions. We", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "score": 1.0, + "content": "follow [26] to do both captioning and filtering on each frame. However, as mentioned in Section 1,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "videos contain rich semantics and temporal contents at multiple granularities. It is not enough to", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "generate video-level target text such as video captions solely based on frame captions. Thus, we", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 710, + 470, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 470, + 724 + ], + "score": 1.0, + "content": "further perform visual tokenization for each frame to capture features at a finer granularity.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 623, + 506, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 70, + 500, + 326 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 70, + 500, + 326 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 70, + 500, + 326 + ], + "spans": [ + { + "bbox": [ + 110, + 70, + 500, + 326 + ], + "score": 0.976, + "type": "image", + "image_path": "c144de0cb595e241110b349a7da85b61029ef363498acc4c08b0c7647b08efae.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 70, + 500, + 155.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 155.33333333333331, + 500, + 240.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 240.66666666666663, + 500, + 325.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 335, + 506, + 423 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 335, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 505, + 347 + ], + "score": 1.0, + "content": "Figure 2: Overview of VidIL framework. We represent a video in a unified textural representation", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 346, + 504, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 504, + 357 + ], + "score": 1.0, + "content": "containing three semantic levels: visual token level, frame level, and video level. At visual token", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 357, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 368 + ], + "score": 1.0, + "content": "level, we extract salient objects, events, attributes for each sampled frame. At frame level, we perform", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 367, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 380 + ], + "score": 1.0, + "content": "image captioning and filtering. At video level, we construct video representation by aggregating", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 379, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 506, + 390 + ], + "score": 1.0, + "content": "the visual tokens, frame captions and other text modalities such as ASR, using a few-shot temporal-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "aware prompt. We then feed the prompt to a pre-trained language model together with task-specific", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 401, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 413 + ], + "score": 1.0, + "content": "instructions to generate target text for a variety of video-language tasks. Examples of the full prompt", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 411, + 301, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 301, + 424 + ], + "score": 1.0, + "content": "for different tasks can be found in Appendix ??.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6.5 + } + ], + "index": 3.75 + }, + { + "type": "title", + "bbox": [ + 107, + 445, + 375, + 457 + ], + "lines": [ + { + "bbox": [ + 105, + 444, + 376, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 376, + 459 + ], + "score": 1.0, + "content": "3.2 Visual Token Level: Structure-Aware Visual Tokenization", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 466, + 505, + 576 + ], + "lines": [ + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "score": 1.0, + "content": "At this level, we aim to extract the textual representations of salient visual token types, such as", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 477, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 506, + 489 + ], + "score": 1.0, + "content": "objects, events and attributes. We found that pre-defined classes for classification, such as those in", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 487, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 501 + ], + "score": 1.0, + "content": "ImageNet [10], are far from enough for covering the rich semantics in open-domain videos. Thus,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 498, + 507, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 507, + 511 + ], + "score": 1.0, + "content": "instead of using classification-based methods for visual tokenization as in previous work [32, 63],", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 509, + 507, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 507, + 523 + ], + "score": 1.0, + "content": "we adopt a retrieval-based visual tokenization approach by leveraging pre-trained contrastive image-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 521, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 521, + 506, + 532 + ], + "score": 1.0, + "content": "language models. Given a visual token vocabulary which contains all candidate object, event, and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 531, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 544 + ], + "score": 1.0, + "content": "attribute text phrases, we compute the image embedding of a frame and the text embeddings of the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "candidate visual tokens using a contrastive multi-modal encoder, CLIP [44]. We then select top 5", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 554, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 505, + 566 + ], + "score": 1.0, + "content": "visual tokens per frame based on the cosine similarity of the image and text embeddings. An example", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 563, + 396, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 396, + 577 + ], + "score": 1.0, + "content": "of the extracted object tokens can be found in the green part of Figure 2.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 581, + 505, + 701 + ], + "lines": [ + { + "bbox": [ + 105, + 580, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 506, + 593 + ], + "score": 1.0, + "content": "Unlike in images where objects and attributes already cover most visual features, events are more", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "informative in videos. In order to discover events from video frames, we construct our own event", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 602, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 602, + 506, + 615 + ], + "score": 1.0, + "content": "vocabulary by extracting event structures from Visual Genome [19] synsets3 using Semantic Role", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 614, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 506, + 626 + ], + "score": 1.0, + "content": "Labeling. Specifically, we first select the phrases that contain at least one verb and one argument", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "score": 1.0, + "content": "as events. Then we remove highly similar events based on their sentence similarity using Sentence-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 635, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 442, + 649 + ], + "score": 1.0, + "content": "BERT [47] embeddings. For object vocabulary, we adopt OpenImage [20] full classes", + "type": "text" + }, + { + "bbox": [ + 442, + 636, + 471, + 646 + ], + "score": 0.86, + "content": "( { \\sim } 2 0 \\mathbf { k } )", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 635, + 506, + 649 + ], + "score": 1.0, + "content": ", instead", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 644, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 262, + 659 + ], + "score": 1.0, + "content": "of using the visually groundable subset", + "type": "text" + }, + { + "bbox": [ + 262, + 646, + 291, + 657 + ], + "score": 0.83, + "content": "( \\sim 6 0 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 644, + 506, + 659 + ], + "score": 1.0, + "content": "as in concurrent work [69]. We found that using large", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 657, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 506, + 669 + ], + "score": 1.0, + "content": "but noisy vocabulary is more effective than using small but clean vocabulary in our retrieval-based", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 668, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 506, + 680 + ], + "score": 1.0, + "content": "setting with CLIP. For attribute vocabulary, we adopt visual genome attribute synset. In Section 4.6,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 679, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 505, + 691 + ], + "score": 1.0, + "content": "we provide ablation study on the impact of different types of visual tokens. The statistics of visual", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 690, + 323, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 323, + 702 + ], + "score": 1.0, + "content": "token vocabulary can be found in Appendix Table ??.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 27 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 115, + 712, + 481, + 722 + ], + "lines": [ + { + "bbox": [ + 118, + 708, + 482, + 725 + ], + "spans": [ + { + "bbox": [ + 118, + 708, + 482, + 725 + ], + "score": 1.0, + "content": "3We use the keys in Visual Genome [19] object synsets which contains frequent pairs.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 11, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 70, + 500, + 326 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 70, + 500, + 326 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 70, + 500, + 326 + ], + "spans": [ + { + "bbox": [ + 110, + 70, + 500, + 326 + ], + "score": 0.976, + "type": "image", + "image_path": "c144de0cb595e241110b349a7da85b61029ef363498acc4c08b0c7647b08efae.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 70, + 500, + 155.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 155.33333333333331, + 500, + 240.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 240.66666666666663, + 500, + 325.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 335, + 506, + 423 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 335, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 505, + 347 + ], + "score": 1.0, + "content": "Figure 2: Overview of VidIL framework. We represent a video in a unified textural representation", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 346, + 504, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 504, + 357 + ], + "score": 1.0, + "content": "containing three semantic levels: visual token level, frame level, and video level. At visual token", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 357, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 368 + ], + "score": 1.0, + "content": "level, we extract salient objects, events, attributes for each sampled frame. At frame level, we perform", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 367, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 380 + ], + "score": 1.0, + "content": "image captioning and filtering. At video level, we construct video representation by aggregating", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 379, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 506, + 390 + ], + "score": 1.0, + "content": "the visual tokens, frame captions and other text modalities such as ASR, using a few-shot temporal-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "aware prompt. We then feed the prompt to a pre-trained language model together with task-specific", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 401, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 413 + ], + "score": 1.0, + "content": "instructions to generate target text for a variety of video-language tasks. Examples of the full prompt", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 411, + 301, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 301, + 424 + ], + "score": 1.0, + "content": "for different tasks can be found in Appendix ??.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 6.5 + } + ], + "index": 3.75 + }, + { + "type": "title", + "bbox": [ + 107, + 445, + 375, + 457 + ], + "lines": [ + { + "bbox": [ + 105, + 444, + 376, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 376, + 459 + ], + "score": 1.0, + "content": "3.2 Visual Token Level: Structure-Aware Visual Tokenization", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 466, + 505, + 576 + ], + "lines": [ + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 505, + 478 + ], + "score": 1.0, + "content": "At this level, we aim to extract the textual representations of salient visual token types, such as", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 477, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 506, + 489 + ], + "score": 1.0, + "content": "objects, events and attributes. We found that pre-defined classes for classification, such as those in", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 487, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 501 + ], + "score": 1.0, + "content": "ImageNet [10], are far from enough for covering the rich semantics in open-domain videos. Thus,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 498, + 507, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 507, + 511 + ], + "score": 1.0, + "content": "instead of using classification-based methods for visual tokenization as in previous work [32, 63],", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 509, + 507, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 507, + 523 + ], + "score": 1.0, + "content": "we adopt a retrieval-based visual tokenization approach by leveraging pre-trained contrastive image-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 521, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 521, + 506, + 532 + ], + "score": 1.0, + "content": "language models. Given a visual token vocabulary which contains all candidate object, event, and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 531, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 544 + ], + "score": 1.0, + "content": "attribute text phrases, we compute the image embedding of a frame and the text embeddings of the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "candidate visual tokens using a contrastive multi-modal encoder, CLIP [44]. We then select top 5", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 554, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 505, + 566 + ], + "score": 1.0, + "content": "visual tokens per frame based on the cosine similarity of the image and text embeddings. An example", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 563, + 396, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 396, + 577 + ], + "score": 1.0, + "content": "of the extracted object tokens can be found in the green part of Figure 2.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 465, + 507, + 577 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 581, + 505, + 701 + ], + "lines": [ + { + "bbox": [ + 105, + 580, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 506, + 593 + ], + "score": 1.0, + "content": "Unlike in images where objects and attributes already cover most visual features, events are more", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "informative in videos. In order to discover events from video frames, we construct our own event", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 602, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 602, + 506, + 615 + ], + "score": 1.0, + "content": "vocabulary by extracting event structures from Visual Genome [19] synsets3 using Semantic Role", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 614, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 506, + 626 + ], + "score": 1.0, + "content": "Labeling. Specifically, we first select the phrases that contain at least one verb and one argument", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "score": 1.0, + "content": "as events. Then we remove highly similar events based on their sentence similarity using Sentence-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 635, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 442, + 649 + ], + "score": 1.0, + "content": "BERT [47] embeddings. For object vocabulary, we adopt OpenImage [20] full classes", + "type": "text" + }, + { + "bbox": [ + 442, + 636, + 471, + 646 + ], + "score": 0.86, + "content": "( { \\sim } 2 0 \\mathbf { k } )", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 635, + 506, + 649 + ], + "score": 1.0, + "content": ", instead", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 644, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 262, + 659 + ], + "score": 1.0, + "content": "of using the visually groundable subset", + "type": "text" + }, + { + "bbox": [ + 262, + 646, + 291, + 657 + ], + "score": 0.83, + "content": "( \\sim 6 0 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 644, + 506, + 659 + ], + "score": 1.0, + "content": "as in concurrent work [69]. We found that using large", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 657, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 506, + 669 + ], + "score": 1.0, + "content": "but noisy vocabulary is more effective than using small but clean vocabulary in our retrieval-based", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 668, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 506, + 680 + ], + "score": 1.0, + "content": "setting with CLIP. For attribute vocabulary, we adopt visual genome attribute synset. In Section 4.6,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 679, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 505, + 691 + ], + "score": 1.0, + "content": "we provide ablation study on the impact of different types of visual tokens. The statistics of visual", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 690, + 323, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 323, + 702 + ], + "score": 1.0, + "content": "token vocabulary can be found in Appendix Table ??.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 580, + 506, + 702 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 74, + 479, + 183 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 74, + 479, + 183 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 74, + 479, + 183 + ], + "spans": [ + { + "bbox": [ + 108, + 74, + 479, + 183 + ], + "score": 0.97, + "type": "image", + "image_path": "331592c36e0944e6b5f6973236f3089e33d229d668d68f0224a0f67fdb784fba.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 74, + 479, + 110.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 110.33333333333334, + 479, + 146.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 146.66666666666669, + 479, + 183.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 103, + 192, + 505, + 214 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 192, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 505, + 204 + ], + "score": 1.0, + "content": "Figure 3: Temporal-aware prompt successfully distinguishes the Sunset and Sunrise scenarios based", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 203, + 477, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 477, + 216 + ], + "score": 1.0, + "content": "on the temporal ordering change of objects and frame captions, while the static prompt fails.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "title", + "bbox": [ + 107, + 225, + 348, + 237 + ], + "lines": [ + { + "bbox": [ + 104, + 222, + 349, + 240 + ], + "spans": [ + { + "bbox": [ + 104, + 222, + 349, + 240 + ], + "score": 1.0, + "content": "3.3 Video Level: Temporal-Aware Few-shot Prompting", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 245, + 505, + 334 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 506, + 258 + ], + "score": 1.0, + "content": "Once we obtain the textual representation from frame level and visual token level, the final step is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 257, + 506, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 506, + 268 + ], + "score": 1.0, + "content": "to put the pieces together to generate a video level target text. The goal is to build a model that", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 268, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 506, + 280 + ], + "score": 1.0, + "content": "can be quickly adapted to any video-to-text generation task with only a few examples. To this", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 279, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 506, + 291 + ], + "score": 1.0, + "content": "end, we propose to leverage large-scale pre-trained language models, such as GPT-3 [6], with a", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 290, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 506, + 302 + ], + "score": 1.0, + "content": "temporal-aware few-shot prompt. As shown in Figure 2, our framework can be readily applied to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 300, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 506, + 313 + ], + "score": 1.0, + "content": "various video-to-text generation tasks, such as video captioning and video question answering, with a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 311, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 506, + 325 + ], + "score": 1.0, + "content": "shared prompt template. The proposed prompting strategy enables a language model to attend to the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 321, + 441, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 441, + 336 + ], + "score": 1.0, + "content": "lower level visual information as well as taking into account the temporal ordering.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 339, + 506, + 481 + ], + "lines": [ + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "score": 1.0, + "content": "Here, we use the video captioning task depicted in Figure 2 to illustrate the details. The few-shot", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 349, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 505, + 362 + ], + "score": 1.0, + "content": "prompt consists of three parts: instruction, few-shot context, and task query. The instruction", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 361, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 506, + 374 + ], + "score": 1.0, + "content": "is a concise description of the generation task, e.g., \"Generate a video caption based on", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 371, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 506, + 383 + ], + "score": 1.0, + "content": "the objects, events, attributes and frame captions. Example:\", which is proved", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 382, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 106, + 382, + 506, + 395 + ], + "score": 1.0, + "content": "to be effective in zero-shot and few-shot settings [6, 59]. The few-shot context contains the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 393, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 506, + 406 + ], + "score": 1.0, + "content": "selected in-context examples as well as the test video instance. Each video instance is repre-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 403, + 501, + 418 + ], + "spans": [ + { + "bbox": [ + 104, + 403, + 501, + 418 + ], + "score": 1.0, + "content": "sented by the aggregated visual tokens4, e.g., \"Objects: First, bath toy. Then,...\"", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 415, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 505, + 428 + ], + "score": 1.0, + "content": "the frame captions, such as \"Frame Captions: First, a toddler playing in a bathtub", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 425, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 506, + 439 + ], + "score": 1.0, + "content": "filled with toys. Then,...\", and the ASR inputs if available, e.g., \"Subtitle:\". Finally, the task query is a task-specific suffix indicating the target text format,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 447, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 104, + 447, + 506, + 461 + ], + "score": 1.0, + "content": "e.g. \"Video Caption:\". For in-context examples (omitted here for simplicity), the task query is", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 457, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 104, + 457, + 506, + 472 + ], + "score": 1.0, + "content": "followed by ground truth annotation, while for the test instance, the generation starts at the end of the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 468, + 153, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 153, + 484 + ], + "score": 1.0, + "content": "task query.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 486, + 504, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 274, + 499 + ], + "score": 1.0, + "content": "Formally, we denote the instruction line as", + "type": "text" + }, + { + "bbox": [ + 274, + 487, + 280, + 496 + ], + "score": 0.27, + "content": "\\mathbf { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 485, + 505, + 499 + ], + "score": 1.0, + "content": ", few-shot context as c, the task query as q, and the target", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 496, + 503, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 134, + 510 + ], + "score": 1.0, + "content": "text as", + "type": "text" + }, + { + "bbox": [ + 134, + 499, + 142, + 509 + ], + "score": 0.42, + "content": "\\mathbf { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 496, + 172, + 510 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 172, + 497, + 252, + 509 + ], + "score": 0.94, + "content": "\\mathbf { y } = ( y _ { 1 } , y _ { 2 } , . . . , y _ { L } )", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 496, + 413, + 510 + ], + "score": 1.0, + "content": ". The generation of the next target token", + "type": "text" + }, + { + "bbox": [ + 414, + 499, + 423, + 509 + ], + "score": 0.81, + "content": "y _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 496, + 503, + 510 + ], + "score": 1.0, + "content": "can be modeled as:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "interline_equation", + "bbox": [ + 243, + 521, + 367, + 542 + ], + "lines": [ + { + "bbox": [ + 243, + 521, + 367, + 542 + ], + "spans": [ + { + "bbox": [ + 243, + 521, + 367, + 542 + ], + "score": 0.93, + "content": "y _ { l } = \\mathop { \\arg \\operatorname* { m a x } } _ { y } p ( y | \\mathbf { s } , \\mathbf { c } , \\mathbf { q } , y _ { < l } )", + "type": "interline_equation", + "image_path": "bc7c063980fb4da70b5248c3574a151aaa1f1ceeacd34ba3434960d3c7181041.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 243, + 521, + 367, + 542 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 547, + 505, + 690 + ], + "lines": [ + { + "bbox": [ + 105, + 547, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 505, + 561 + ], + "score": 1.0, + "content": "In order to capture the temporal dynamics between frames and visual tokens, we further propose to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 560, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 505, + 571 + ], + "score": 1.0, + "content": "inject temporal markers to the prompt. As shown in the few-shot context in Figure 2, each visual", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 568, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 104, + 568, + 506, + 585 + ], + "score": 1.0, + "content": "token and frame caption is prefixed with a natural language phrase indicating its temporal ordering,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 580, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 104, + 580, + 506, + 594 + ], + "score": 1.0, + "content": "e.g., \"First,\",\"Then,\", and \"Finally,\". We found adding the temporal marker can make the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 592, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 506, + 604 + ], + "score": 1.0, + "content": "language model conditioned on not only literal but also temporal information of the context. We show", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 604, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 615 + ], + "score": 1.0, + "content": "an example in Figure 3, where we compare our temporal-aware prompt with a static prompt on video", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 614, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 506, + 626 + ], + "score": 1.0, + "content": "captioning using InstructGPT. Again, the in-context examples are omitted here, which can be found", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "score": 1.0, + "content": "in Appendix ??. In this example, the only difference between these two contexts is the ordering of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 635, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 506, + 649 + ], + "score": 1.0, + "content": "the visual tokens and the frame captions. For the context on the left, where \"sun moving\" appears", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 645, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 660 + ], + "score": 1.0, + "content": "before \"night sky\", we are expected to see a story talking about sunset, while for the context on", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "the right, we are expected to see sunrise. We can see the static prompt generates captions about", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "sunset for both contexts, while the temporal-aware prompt can capture temporal ordering correctly", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 679, + 301, + 692 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 301, + 692 + ], + "score": 1.0, + "content": "and generate sunrise for the context on the right.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 36 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 701, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 118, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 118, + 699, + 506, + 714 + ], + "score": 1.0, + "content": "4To obtain video level visual tokens, the visual tokens extracted from each frame are further ranked and", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 711, + 428, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 428, + 722 + ], + "score": 1.0, + "content": "ordered based on frequency and frame index. More details can be found in Appendix ??.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 74, + 479, + 183 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 74, + 479, + 183 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 74, + 479, + 183 + ], + "spans": [ + { + "bbox": [ + 108, + 74, + 479, + 183 + ], + "score": 0.97, + "type": "image", + "image_path": "331592c36e0944e6b5f6973236f3089e33d229d668d68f0224a0f67fdb784fba.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 74, + 479, + 110.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 110.33333333333334, + 479, + 146.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 146.66666666666669, + 479, + 183.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 103, + 192, + 505, + 214 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 192, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 505, + 204 + ], + "score": 1.0, + "content": "Figure 3: Temporal-aware prompt successfully distinguishes the Sunset and Sunrise scenarios based", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 203, + 477, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 477, + 216 + ], + "score": 1.0, + "content": "on the temporal ordering change of objects and frame captions, while the static prompt fails.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "title", + "bbox": [ + 107, + 225, + 348, + 237 + ], + "lines": [ + { + "bbox": [ + 104, + 222, + 349, + 240 + ], + "spans": [ + { + "bbox": [ + 104, + 222, + 349, + 240 + ], + "score": 1.0, + "content": "3.3 Video Level: Temporal-Aware Few-shot Prompting", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 245, + 505, + 334 + ], + "lines": [ + { + "bbox": [ + 106, + 245, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 506, + 258 + ], + "score": 1.0, + "content": "Once we obtain the textual representation from frame level and visual token level, the final step is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 257, + 506, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 506, + 268 + ], + "score": 1.0, + "content": "to put the pieces together to generate a video level target text. The goal is to build a model that", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 268, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 506, + 280 + ], + "score": 1.0, + "content": "can be quickly adapted to any video-to-text generation task with only a few examples. To this", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 279, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 506, + 291 + ], + "score": 1.0, + "content": "end, we propose to leverage large-scale pre-trained language models, such as GPT-3 [6], with a", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 290, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 506, + 302 + ], + "score": 1.0, + "content": "temporal-aware few-shot prompt. As shown in Figure 2, our framework can be readily applied to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 300, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 506, + 313 + ], + "score": 1.0, + "content": "various video-to-text generation tasks, such as video captioning and video question answering, with a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 311, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 506, + 325 + ], + "score": 1.0, + "content": "shared prompt template. The proposed prompting strategy enables a language model to attend to the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 321, + 441, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 441, + 336 + ], + "score": 1.0, + "content": "lower level visual information as well as taking into account the temporal ordering.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 245, + 506, + 336 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 339, + 506, + 481 + ], + "lines": [ + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 506, + 351 + ], + "score": 1.0, + "content": "Here, we use the video captioning task depicted in Figure 2 to illustrate the details. The few-shot", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 349, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 505, + 362 + ], + "score": 1.0, + "content": "prompt consists of three parts: instruction, few-shot context, and task query. The instruction", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 361, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 506, + 374 + ], + "score": 1.0, + "content": "is a concise description of the generation task, e.g., \"Generate a video caption based on", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 371, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 506, + 383 + ], + "score": 1.0, + "content": "the objects, events, attributes and frame captions. Example:\", which is proved", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 382, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 106, + 382, + 506, + 395 + ], + "score": 1.0, + "content": "to be effective in zero-shot and few-shot settings [6, 59]. The few-shot context contains the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 393, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 506, + 406 + ], + "score": 1.0, + "content": "selected in-context examples as well as the test video instance. Each video instance is repre-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 403, + 501, + 418 + ], + "spans": [ + { + "bbox": [ + 104, + 403, + 501, + 418 + ], + "score": 1.0, + "content": "sented by the aggregated visual tokens4, e.g., \"Objects: First, bath toy. Then,...\"", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 415, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 505, + 428 + ], + "score": 1.0, + "content": "the frame captions, such as \"Frame Captions: First, a toddler playing in a bathtub", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 425, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 506, + 439 + ], + "score": 1.0, + "content": "filled with toys. Then,...\", and the ASR inputs if available, e.g., \"Subtitle:\". Finally, the task query is a task-specific suffix indicating the target text format,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 447, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 104, + 447, + 506, + 461 + ], + "score": 1.0, + "content": "e.g. \"Video Caption:\". For in-context examples (omitted here for simplicity), the task query is", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 457, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 104, + 457, + 506, + 472 + ], + "score": 1.0, + "content": "followed by ground truth annotation, while for the test instance, the generation starts at the end of the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 468, + 153, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 153, + 484 + ], + "score": 1.0, + "content": "task query.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 20, + "bbox_fs": [ + 104, + 338, + 506, + 484 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 486, + 504, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 274, + 499 + ], + "score": 1.0, + "content": "Formally, we denote the instruction line as", + "type": "text" + }, + { + "bbox": [ + 274, + 487, + 280, + 496 + ], + "score": 0.27, + "content": "\\mathbf { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 485, + 505, + 499 + ], + "score": 1.0, + "content": ", few-shot context as c, the task query as q, and the target", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 496, + 503, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 134, + 510 + ], + "score": 1.0, + "content": "text as", + "type": "text" + }, + { + "bbox": [ + 134, + 499, + 142, + 509 + ], + "score": 0.42, + "content": "\\mathbf { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 496, + 172, + 510 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 172, + 497, + 252, + 509 + ], + "score": 0.94, + "content": "\\mathbf { y } = ( y _ { 1 } , y _ { 2 } , . . . , y _ { L } )", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 496, + 413, + 510 + ], + "score": 1.0, + "content": ". The generation of the next target token", + "type": "text" + }, + { + "bbox": [ + 414, + 499, + 423, + 509 + ], + "score": 0.81, + "content": "y _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 496, + 503, + 510 + ], + "score": 1.0, + "content": "can be modeled as:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 485, + 505, + 510 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 243, + 521, + 367, + 542 + ], + "lines": [ + { + "bbox": [ + 243, + 521, + 367, + 542 + ], + "spans": [ + { + "bbox": [ + 243, + 521, + 367, + 542 + ], + "score": 0.93, + "content": "y _ { l } = \\mathop { \\arg \\operatorname* { m a x } } _ { y } p ( y | \\mathbf { s } , \\mathbf { c } , \\mathbf { q } , y _ { < l } )", + "type": "interline_equation", + "image_path": "bc7c063980fb4da70b5248c3574a151aaa1f1ceeacd34ba3434960d3c7181041.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 243, + 521, + 367, + 542 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 547, + 505, + 690 + ], + "lines": [ + { + "bbox": [ + 105, + 547, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 505, + 561 + ], + "score": 1.0, + "content": "In order to capture the temporal dynamics between frames and visual tokens, we further propose to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 560, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 505, + 571 + ], + "score": 1.0, + "content": "inject temporal markers to the prompt. As shown in the few-shot context in Figure 2, each visual", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 568, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 104, + 568, + 506, + 585 + ], + "score": 1.0, + "content": "token and frame caption is prefixed with a natural language phrase indicating its temporal ordering,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 580, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 104, + 580, + 506, + 594 + ], + "score": 1.0, + "content": "e.g., \"First,\",\"Then,\", and \"Finally,\". We found adding the temporal marker can make the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 592, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 506, + 604 + ], + "score": 1.0, + "content": "language model conditioned on not only literal but also temporal information of the context. We show", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 604, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 615 + ], + "score": 1.0, + "content": "an example in Figure 3, where we compare our temporal-aware prompt with a static prompt on video", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 614, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 506, + 626 + ], + "score": 1.0, + "content": "captioning using InstructGPT. Again, the in-context examples are omitted here, which can be found", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "score": 1.0, + "content": "in Appendix ??. In this example, the only difference between these two contexts is the ordering of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 635, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 506, + 649 + ], + "score": 1.0, + "content": "the visual tokens and the frame captions. For the context on the left, where \"sun moving\" appears", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 645, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 660 + ], + "score": 1.0, + "content": "before \"night sky\", we are expected to see a story talking about sunset, while for the context on", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "the right, we are expected to see sunrise. We can see the static prompt generates captions about", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "sunset for both contexts, while the temporal-aware prompt can capture temporal ordering correctly", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 679, + 301, + 692 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 301, + 692 + ], + "score": 1.0, + "content": "and generate sunrise for the context on the right.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 36, + "bbox_fs": [ + 104, + 547, + 506, + 692 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 71, + 191, + 85 + ], + "lines": [ + { + "bbox": [ + 104, + 69, + 193, + 88 + ], + "spans": [ + { + "bbox": [ + 104, + 69, + 193, + 88 + ], + "score": 1.0, + "content": "4 Experiments", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 96, + 215, + 108 + ], + "lines": [ + { + "bbox": [ + 104, + 93, + 217, + 112 + ], + "spans": [ + { + "bbox": [ + 104, + 93, + 217, + 112 + ], + "score": 1.0, + "content": "4.1 Experimental Setup", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 117, + 505, + 171 + ], + "lines": [ + { + "bbox": [ + 106, + 115, + 505, + 130 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 505, + 130 + ], + "score": 1.0, + "content": "To comprehensively evaluate our model, we show results on four video-language understanding", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 127, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 505, + 141 + ], + "score": 1.0, + "content": "tasks in few-shot settings: video captioning, video question answering (QA), video-language event", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 139, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 139, + 505, + 151 + ], + "score": 1.0, + "content": "prediction, and text-video retrieval. We compare our approach with state-of-the-art approaches on", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 150, + 504, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 504, + 162 + ], + "score": 1.0, + "content": "five benchmarks, i.e, MSR-VTT [62], MSVD [7], VaTeX [57], YouCook2 [74], and VLEP [23]. The", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 160, + 485, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 485, + 172 + ], + "score": 1.0, + "content": "statistics of the datasets can be found in Table 1. For more details please refer to Appendix ??.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 179, + 297, + 288 + ], + "lines": [ + { + "bbox": [ + 106, + 179, + 297, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 179, + 297, + 190 + ], + "score": 1.0, + "content": "Implementation Details. We use CLIP-L/146", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 190, + 298, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 298, + 202 + ], + "score": 1.0, + "content": "as our default encoder for visual tokenization.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 201, + 298, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 298, + 213 + ], + "score": 1.0, + "content": "We adopt BLIP captioning checkpoint7 fine-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 211, + 297, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 297, + 225 + ], + "score": 1.0, + "content": "tuned on COCO [31] for frame captioning. We", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 222, + 298, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 298, + 236 + ], + "score": 1.0, + "content": "use InstructGPT [40] as our default language", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 234, + 297, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 297, + 246 + ], + "score": 1.0, + "content": "model for generating text conditioned on the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 245, + 298, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 298, + 257 + ], + "score": 1.0, + "content": "few-shot prompt. To construct event vocabulary,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 255, + 297, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 297, + 267 + ], + "score": 1.0, + "content": "we use the semantic role labeling model from", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 266, + 297, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 297, + 279 + ], + "score": 1.0, + "content": "AllenNLP8. The experiments are conducted on", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 277, + 298, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 298, + 289 + ], + "score": 1.0, + "content": "2 NVIDIA V100 (16GB) GPUs. All few-shot", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 11.5 + }, + { + "type": "table", + "bbox": [ + 304, + 195, + 504, + 279 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 307, + 178, + 502, + 190 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 306, + 177, + 502, + 191 + ], + "spans": [ + { + "bbox": [ + 306, + 177, + 502, + 191 + ], + "score": 1.0, + "content": "Table 1: Statistics of datasets in our experiments", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "table_body", + "bbox": [ + 304, + 195, + 504, + 279 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 195, + 504, + 279 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 504, + 279 + ], + "score": 0.976, + "html": "
DatasetTaskSplit Count # train /#eval
MSR-VTT[62]Captioning; QA6,513 /2.990
MSR-VTT[62]Retrieval7,010 / 1,000
MSVD [7]Question Answering30,933 /13,157
VaTeX v1.15 [57]Captioning; Retrieval25,991/6.000
YouCook2[74]Captioning10,337 /3,492
VLEP [23]Event Prediction20,142/4,192
", + "type": "table", + "image_path": "92dd6fadc5a80d33d4e99b29ba373e80a80afa39fcabf656f24d4c71b6eab48f.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 304, + 195, + 504, + 209.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 304, + 209.0, + 504, + 223.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 304, + 223.0, + 504, + 237.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 304, + 237.0, + 504, + 251.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 304, + 251.0, + 504, + 265.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 304, + 265.0, + 504, + 279.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 107, + 289, + 498, + 300 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 288, + 500, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 500, + 301 + ], + "score": 1.0, + "content": "finetuning on baselines and semi-supervised training are performed on 2 Nvidia V100 16G GPUs.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 313, + 505, + 444 + ], + "lines": [ + { + "bbox": [ + 105, + 311, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 505, + 325 + ], + "score": 1.0, + "content": "In-context Example Selection. From our preliminary experiments, we find that the generation", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "score": 1.0, + "content": "performance is sensitive to the quality of in-context examples. For example, for QA tasks such", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "score": 1.0, + "content": "as MSVD-QA where the annotations are automatically generated, the pair in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "score": 1.0, + "content": "randomly selected in-context examples can be only weakly-correlated with the video context. Thus,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 357, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 506, + 369 + ], + "score": 1.0, + "content": "instead of using a fixed prompt for each query, we dynamically filter out the irrelevant in-context", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 368, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 305, + 380 + ], + "score": 1.0, + "content": "examples. Specifically, given a randomly sampled", + "type": "text" + }, + { + "bbox": [ + 305, + 368, + 316, + 378 + ], + "score": 0.57, + "content": "M .", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 368, + 506, + 380 + ], + "score": 1.0, + "content": "-shot support set from the training set, we select", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 378, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 151, + 390 + ], + "score": 1.0, + "content": "a subset of", + "type": "text" + }, + { + "bbox": [ + 151, + 379, + 160, + 388 + ], + "score": 0.81, + "content": "N _ { ☉ }", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 378, + 505, + 390 + ], + "score": 1.0, + "content": "-shots as in-context examples based on their SentenceBERT [47] similarities with text", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "queries. Furthermore, we reorder the selected examples in ascending order based on the similarity", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 401, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 505, + 413 + ], + "score": 1.0, + "content": "score to account for the recency bias [72] in large language models. For QA tasks, we choose the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "most relevant in-context examples by comparing with questions. While for captioning task, we", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 422, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 361, + 434 + ], + "score": 1.0, + "content": "compare with frame captions. If not otherwise specified, we use", + "type": "text" + }, + { + "bbox": [ + 361, + 422, + 388, + 432 + ], + "score": 0.88, + "content": "M { = } I O", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 422, + 405, + 434 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 405, + 422, + 425, + 432 + ], + "score": 0.8, + "content": "N { = } 5", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 422, + 506, + 434 + ], + "score": 1.0, + "content": ", which we consider", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 432, + 186, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 186, + 446 + ], + "score": 1.0, + "content": "as 10-shot training.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 30.5 + }, + { + "type": "title", + "bbox": [ + 108, + 457, + 246, + 469 + ], + "lines": [ + { + "bbox": [ + 105, + 456, + 248, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 248, + 473 + ], + "score": 1.0, + "content": "4.2 Few-shot Video Captioning", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 478, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 106, + 478, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 506, + 491 + ], + "score": 1.0, + "content": "We report BLEU-4 [41], ROUGE-L [30], METEOR [5], and CIDEr [54] scores on three video caption-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 489, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 506, + 502 + ], + "score": 1.0, + "content": "ing benchmarks covering both open-domain (MSR-VTT, VaTeX) and domain-specific (YouCook2)", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 499, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 506, + 513 + ], + "score": 1.0, + "content": "videos. We compare with both state-of-the-art video captioner (UniVL [36]) and image captioner", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "(BLIP [26]). In order to implement the BLIP baseline for few-shot video captioning, we extend the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 522, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 506, + 536 + ], + "score": 1.0, + "content": "approach used for text-video retrieval evaluation in [26] to video-language training. Specifically,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "we concatenate the visual features of sampled frames and then feed them into the image-grounded", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 544, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 506, + 556 + ], + "score": 1.0, + "content": "text-encoder to compute the language modeling loss. This is equivalent to stitching the sampled", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "frames into a large image and then feeding it to BLIP for image captioning. We found that this simple", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 566, + 272, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 272, + 578 + ], + "score": 1.0, + "content": "approach results in very strong baselines.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 106, + 582, + 505, + 670 + ], + "lines": [ + { + "bbox": [ + 106, + 582, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 506, + 594 + ], + "score": 1.0, + "content": "As shown in Table 2, existing methods have strong bias on certain datasets. For example, UniVL", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 593, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 506, + 605 + ], + "score": 1.0, + "content": "performs well on YouCook2 but fails on MSR-VTT and VaTeX, while BLIP performs the oppo-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 603, + 506, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 506, + 616 + ], + "score": 1.0, + "content": "site. This is because UniVL is pretrained on HowTo100M which favors instructional videos, i.e.,", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 614, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 506, + 628 + ], + "score": 1.0, + "content": "YouCook2, while BLIP is pre-training on image-caption pairs which favors description-style captions,", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 104, + 625, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 104, + 625, + 505, + 639 + ], + "score": 1.0, + "content": "i.e., MSR-VTT and VaTeX. On the contrary, our model performs competitively on both open-domain", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 637, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 505, + 649 + ], + "score": 1.0, + "content": "and instructional videos, and significantly outperforms the baselines on the average CIDEr score", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 648, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 505, + 660 + ], + "score": 1.0, + "content": "across all three benchmarks. This indicates that by leveraging language models, we can maintain", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 659, + 456, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 456, + 671 + ], + "score": 1.0, + "content": "strong few-shot ability regardless of the video domain or the target caption distribution.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 50.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 680, + 492, + 721 + ], + "lines": [ + { + "bbox": [ + 118, + 678, + 372, + 691 + ], + "spans": [ + { + "bbox": [ + 118, + 678, + 372, + 691 + ], + "score": 1.0, + "content": "6https://huggingface.co/openai/clip-vit-large-patch14", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 689, + 390, + 703 + ], + "spans": [ + { + "bbox": [ + 118, + 689, + 390, + 703 + ], + "score": 1.0, + "content": "7https://github.com/salesforce/BLIP#finetuned-checkpoints", + "type": "text" + } + ] + }, + { + "bbox": [ + 117, + 699, + 492, + 714 + ], + "spans": [ + { + "bbox": [ + 117, + 699, + 492, + 714 + ], + "score": 1.0, + "content": "8https://docs.allennlp.org/models/main/models/structured_prediction/predictors/", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 710, + 129, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 129, + 723 + ], + "score": 1.0, + "content": "srl/", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 742, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 71, + 191, + 85 + ], + "lines": [ + { + "bbox": [ + 104, + 69, + 193, + 88 + ], + "spans": [ + { + "bbox": [ + 104, + 69, + 193, + 88 + ], + "score": 1.0, + "content": "4 Experiments", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 96, + 215, + 108 + ], + "lines": [ + { + "bbox": [ + 104, + 93, + 217, + 112 + ], + "spans": [ + { + "bbox": [ + 104, + 93, + 217, + 112 + ], + "score": 1.0, + "content": "4.1 Experimental Setup", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 117, + 505, + 171 + ], + "lines": [ + { + "bbox": [ + 106, + 115, + 505, + 130 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 505, + 130 + ], + "score": 1.0, + "content": "To comprehensively evaluate our model, we show results on four video-language understanding", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 127, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 505, + 141 + ], + "score": 1.0, + "content": "tasks in few-shot settings: video captioning, video question answering (QA), video-language event", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 139, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 139, + 505, + 151 + ], + "score": 1.0, + "content": "prediction, and text-video retrieval. We compare our approach with state-of-the-art approaches on", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 150, + 504, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 504, + 162 + ], + "score": 1.0, + "content": "five benchmarks, i.e, MSR-VTT [62], MSVD [7], VaTeX [57], YouCook2 [74], and VLEP [23]. The", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 160, + 485, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 485, + 172 + ], + "score": 1.0, + "content": "statistics of the datasets can be found in Table 1. For more details please refer to Appendix ??.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 115, + 505, + 172 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 179, + 297, + 288 + ], + "lines": [ + { + "bbox": [ + 106, + 179, + 297, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 179, + 297, + 190 + ], + "score": 1.0, + "content": "Implementation Details. We use CLIP-L/146", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 190, + 298, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 298, + 202 + ], + "score": 1.0, + "content": "as our default encoder for visual tokenization.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 201, + 298, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 298, + 213 + ], + "score": 1.0, + "content": "We adopt BLIP captioning checkpoint7 fine-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 211, + 297, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 297, + 225 + ], + "score": 1.0, + "content": "tuned on COCO [31] for frame captioning. We", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 222, + 298, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 298, + 236 + ], + "score": 1.0, + "content": "use InstructGPT [40] as our default language", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 234, + 297, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 297, + 246 + ], + "score": 1.0, + "content": "model for generating text conditioned on the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 245, + 298, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 298, + 257 + ], + "score": 1.0, + "content": "few-shot prompt. To construct event vocabulary,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 255, + 297, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 297, + 267 + ], + "score": 1.0, + "content": "we use the semantic role labeling model from", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 266, + 297, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 297, + 279 + ], + "score": 1.0, + "content": "AllenNLP8. The experiments are conducted on", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 277, + 298, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 298, + 289 + ], + "score": 1.0, + "content": "2 NVIDIA V100 (16GB) GPUs. All few-shot", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 179, + 298, + 289 + ] + }, + { + "type": "table", + "bbox": [ + 304, + 195, + 504, + 279 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 307, + 178, + 502, + 190 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 306, + 177, + 502, + 191 + ], + "spans": [ + { + "bbox": [ + 306, + 177, + 502, + 191 + ], + "score": 1.0, + "content": "Table 1: Statistics of datasets in our experiments", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "table_body", + "bbox": [ + 304, + 195, + 504, + 279 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 195, + 504, + 279 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 504, + 279 + ], + "score": 0.976, + "html": "
DatasetTaskSplit Count # train /#eval
MSR-VTT[62]Captioning; QA6,513 /2.990
MSR-VTT[62]Retrieval7,010 / 1,000
MSVD [7]Question Answering30,933 /13,157
VaTeX v1.15 [57]Captioning; Retrieval25,991/6.000
YouCook2[74]Captioning10,337 /3,492
VLEP [23]Event Prediction20,142/4,192
", + "type": "table", + "image_path": "92dd6fadc5a80d33d4e99b29ba373e80a80afa39fcabf656f24d4c71b6eab48f.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 304, + 195, + 504, + 209.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 304, + 209.0, + 504, + 223.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 304, + 223.0, + 504, + 237.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 304, + 237.0, + 504, + 251.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 304, + 251.0, + 504, + 265.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 304, + 265.0, + 504, + 279.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 107, + 289, + 498, + 300 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 288, + 500, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 500, + 301 + ], + "score": 1.0, + "content": "finetuning on baselines and semi-supervised training are performed on 2 Nvidia V100 16G GPUs.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 313, + 505, + 444 + ], + "lines": [ + { + "bbox": [ + 105, + 311, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 505, + 325 + ], + "score": 1.0, + "content": "In-context Example Selection. From our preliminary experiments, we find that the generation", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "score": 1.0, + "content": "performance is sensitive to the quality of in-context examples. For example, for QA tasks such", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "score": 1.0, + "content": "as MSVD-QA where the annotations are automatically generated, the pair in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 358 + ], + "score": 1.0, + "content": "randomly selected in-context examples can be only weakly-correlated with the video context. Thus,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 357, + 506, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 506, + 369 + ], + "score": 1.0, + "content": "instead of using a fixed prompt for each query, we dynamically filter out the irrelevant in-context", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 368, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 305, + 380 + ], + "score": 1.0, + "content": "examples. Specifically, given a randomly sampled", + "type": "text" + }, + { + "bbox": [ + 305, + 368, + 316, + 378 + ], + "score": 0.57, + "content": "M .", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 368, + 506, + 380 + ], + "score": 1.0, + "content": "-shot support set from the training set, we select", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 378, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 151, + 390 + ], + "score": 1.0, + "content": "a subset of", + "type": "text" + }, + { + "bbox": [ + 151, + 379, + 160, + 388 + ], + "score": 0.81, + "content": "N _ { ☉ }", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 378, + 505, + 390 + ], + "score": 1.0, + "content": "-shots as in-context examples based on their SentenceBERT [47] similarities with text", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "queries. Furthermore, we reorder the selected examples in ascending order based on the similarity", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 401, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 505, + 413 + ], + "score": 1.0, + "content": "score to account for the recency bias [72] in large language models. For QA tasks, we choose the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "most relevant in-context examples by comparing with questions. While for captioning task, we", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 422, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 361, + 434 + ], + "score": 1.0, + "content": "compare with frame captions. If not otherwise specified, we use", + "type": "text" + }, + { + "bbox": [ + 361, + 422, + 388, + 432 + ], + "score": 0.88, + "content": "M { = } I O", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 422, + 405, + 434 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 405, + 422, + 425, + 432 + ], + "score": 0.8, + "content": "N { = } 5", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 422, + 506, + 434 + ], + "score": 1.0, + "content": ", which we consider", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 432, + 186, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 186, + 446 + ], + "score": 1.0, + "content": "as 10-shot training.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 311, + 506, + 446 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 457, + 246, + 469 + ], + "lines": [ + { + "bbox": [ + 105, + 456, + 248, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 248, + 473 + ], + "score": 1.0, + "content": "4.2 Few-shot Video Captioning", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 478, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 106, + 478, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 506, + 491 + ], + "score": 1.0, + "content": "We report BLEU-4 [41], ROUGE-L [30], METEOR [5], and CIDEr [54] scores on three video caption-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 489, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 506, + 502 + ], + "score": 1.0, + "content": "ing benchmarks covering both open-domain (MSR-VTT, VaTeX) and domain-specific (YouCook2)", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 499, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 506, + 513 + ], + "score": 1.0, + "content": "videos. We compare with both state-of-the-art video captioner (UniVL [36]) and image captioner", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 505, + 524 + ], + "score": 1.0, + "content": "(BLIP [26]). In order to implement the BLIP baseline for few-shot video captioning, we extend the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 522, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 506, + 536 + ], + "score": 1.0, + "content": "approach used for text-video retrieval evaluation in [26] to video-language training. Specifically,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "we concatenate the visual features of sampled frames and then feed them into the image-grounded", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 544, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 506, + 556 + ], + "score": 1.0, + "content": "text-encoder to compute the language modeling loss. This is equivalent to stitching the sampled", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "frames into a large image and then feeding it to BLIP for image captioning. We found that this simple", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 566, + 272, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 272, + 578 + ], + "score": 1.0, + "content": "approach results in very strong baselines.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 478, + 506, + 578 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 582, + 505, + 670 + ], + "lines": [ + { + "bbox": [ + 106, + 582, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 506, + 594 + ], + "score": 1.0, + "content": "As shown in Table 2, existing methods have strong bias on certain datasets. For example, UniVL", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 593, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 593, + 506, + 605 + ], + "score": 1.0, + "content": "performs well on YouCook2 but fails on MSR-VTT and VaTeX, while BLIP performs the oppo-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 603, + 506, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 506, + 616 + ], + "score": 1.0, + "content": "site. This is because UniVL is pretrained on HowTo100M which favors instructional videos, i.e.,", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 614, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 506, + 628 + ], + "score": 1.0, + "content": "YouCook2, while BLIP is pre-training on image-caption pairs which favors description-style captions,", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 104, + 625, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 104, + 625, + 505, + 639 + ], + "score": 1.0, + "content": "i.e., MSR-VTT and VaTeX. On the contrary, our model performs competitively on both open-domain", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 637, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 505, + 649 + ], + "score": 1.0, + "content": "and instructional videos, and significantly outperforms the baselines on the average CIDEr score", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 648, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 505, + 660 + ], + "score": 1.0, + "content": "across all three benchmarks. This indicates that by leveraging language models, we can maintain", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 659, + 456, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 456, + 671 + ], + "score": 1.0, + "content": "strong few-shot ability regardless of the video domain or the target caption distribution.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 50.5, + "bbox_fs": [ + 104, + 582, + 506, + 671 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 150, + 505, + 312 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 77, + 505, + 144 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 77, + 506, + 92 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 506, + 92 + ], + "score": 1.0, + "content": "Table 2: 10-shot video captioning results. ♠ indicates concurrent work. The reported Flamingo [2]", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 89, + 506, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 89, + 506, + 102 + ], + "score": 1.0, + "content": "results are using 16 shots. #VideoPT represents the number of videos used for pre-training. B-4, R-L,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 99, + 506, + 112 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 116, + 110 + ], + "score": 0.3, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 99, + 120, + 112 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 120, + 100, + 128, + 110 + ], + "score": 0.53, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 99, + 372, + 112 + ], + "score": 1.0, + "content": "represents BLEU-4, ROUGE-L, METEOR and CIDEr. Avg", + "type": "text" + }, + { + "bbox": [ + 373, + 100, + 382, + 110 + ], + "score": 0.29, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 99, + 506, + 112 + ], + "score": 1.0, + "content": "represents the average CIDEr", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "score across all available benchmarks. ASR indicates whether the model has access to the ASR", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 120, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 145, + 135 + ], + "score": 1.0, + "content": "subtitles.", + "type": "text" + }, + { + "bbox": [ + 145, + 122, + 169, + 132 + ], + "score": 0.34, + "content": "B L I P", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 120, + 186, + 135 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 186, + 122, + 221, + 134 + ], + "score": 0.91, + "content": "B L I P _ { c a p }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 120, + 506, + 135 + ], + "score": 1.0, + "content": "use the pretrained checkpoint and the finetuned checkpoint on COCO", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 133, + 352, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 133, + 352, + 145 + ], + "score": 1.0, + "content": "captioning. All results are averaged over three random seeds.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "table_body", + "bbox": [ + 106, + 150, + 505, + 312 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 150, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 505, + 312 + ], + "score": 0.945, + "html": "
Method#VideOpr ASRB-4R-LMMSR-VTTCaptionYouCook2 Caption CVaTex CaptionAvg C
CB-4R-LMB-4 R-L MC
Few-shot
UniVL1.2MNo2.1 22.5 9.53.63.325.3 11.634.11.715.78.02.113.3
BLIP0No27.7 43.0 23.0 39.50.79.0 3.411.513.539.515.420.723.9
BLIPcap0No21.648.022.7 30.23.78.6 3.89.420.741.517.428.922.8
VidIL(ours)0No26.0 51.7 24.7 36.32.622.9 9.527.022.243.620.036.733.3
UniVL1.2MYes-=-4.326.4 12.248.62.717.710.23.426.0
VidIL(ours)0Yes10.7 35.919.4 111.623.2 44.2 20.6 38.975.3
Flamingo-3B(16)27MNo73.257.1
Flamingo-80B(16) 27MNo84.262.8=
Fine-tuning
UniVL1.2MNo42.0 61.0 29.0 50.1[11.2 40.1 17.6 127.0|22.8 38.6 22.3 33.470.2
UniVL1.2MYes=16.6 45.7 21.6176.823.7 39.3 22.7 35.6106.2
", + "type": "table", + "image_path": "bdc0755cadf49e276175a4fa571481c0553518341285949af409281797b235b2.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 106, + 150, + 505, + 204.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 106, + 204.0, + 505, + 258.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 106, + 258.0, + 505, + 312.0 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 4.75 + }, + { + "type": "text", + "bbox": [ + 106, + 333, + 505, + 377 + ], + "lines": [ + { + "bbox": [ + 105, + 333, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 345 + ], + "score": 1.0, + "content": "As discussed in Section 1, video captions describe the content in various semantic levels. The N-gram", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 343, + 507, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 507, + 357 + ], + "score": 1.0, + "content": "based metric may not fairly reflect the models’ performance in capturing the video-caption alignment.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "We further verify this hypothesis in Section 4.5. Thus, in addition to automatic metrics, we include", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 366, + 428, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 428, + 378 + ], + "score": 1.0, + "content": "qualitative examples illustrated in Figure 4. More examples are in Appendix ??.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 382, + 505, + 448 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 506, + 395 + ], + "score": 1.0, + "content": "Additionally, for most existing methods and also concurrent work, e.g., Flamingo [2], adding a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "new modality often requires a dedicated model redesign or retraining. However, the nature of our", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 404, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 505, + 416 + ], + "score": 1.0, + "content": "framework, where we use a unified textual representation for each level, makes it highly flexible for", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 415, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 505, + 426 + ], + "score": 1.0, + "content": "incorporating new modalities. As shown in row 6 in Table, our model can effectively utilize extra", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 425, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 506, + 438 + ], + "score": 1.0, + "content": "information from ASR to obtain significantly better few-shot performance on certain datasets such as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 435, + 153, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 153, + 448 + ], + "score": 1.0, + "content": "YouCook2.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5 + }, + { + "type": "image", + "bbox": [ + 106, + 456, + 504, + 599 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 456, + 504, + 599 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 456, + 504, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 504, + 599 + ], + "score": 0.577, + "type": "image", + "image_path": "42488e5e146be037d7c57abcc48a78f8af1cf75e6517bc50d3e1b32bfd715b6a.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 106, + 456, + 504, + 503.6666666666667 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 106, + 503.6666666666667, + 504, + 551.3333333333334 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 106, + 551.3333333333334, + 504, + 599.0 + ], + "spans": [], + "index": 21 + } + ] + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 609, + 506, + 654 + ], + "lines": [ + { + "bbox": [ + 105, + 609, + 507, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 507, + 622 + ], + "score": 1.0, + "content": "Figure 4: Qualitative examples on video captioning. Grey boxes contain part of the video repre-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 619, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 633 + ], + "score": 1.0, + "content": "sentation from our model. Blue boxes contain caption generation from different models. Green", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 631, + 506, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 506, + 643 + ], + "score": 1.0, + "content": "boxes contain ground truth annotations. Bold green text highlights the correct information that is not", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 642, + 487, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 487, + 655 + ], + "score": 1.0, + "content": "captured in baseline outputs which can be reasoned from our visual tokens and frame captions.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "title", + "bbox": [ + 108, + 668, + 285, + 681 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 287, + 683 + ], + "score": 1.0, + "content": "4.3 Few-shot Video Question Answering", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 689, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 422, + 702 + ], + "score": 1.0, + "content": "We compare the test accuracy of our approach with few-shot pretrained BLIP,", + "type": "text" + }, + { + "bbox": [ + 423, + 689, + 465, + 701 + ], + "score": 0.87, + "content": "\\mathsf { B L I P } _ { V Q A }", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "[26], and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "concurrent work Flamingo [2] on two video question answering benchmarks, MSR-VTT_QA and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 159, + 723 + ], + "score": 1.0, + "content": "MSVD_QA.", + "type": "text" + }, + { + "bbox": [ + 160, + 711, + 201, + 723 + ], + "score": 0.76, + "content": "\\mathsf { B L I P } _ { V Q A }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "represents finetuned BLIP on VQA [4] dataset, which is the previous SOTA", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 150, + 505, + 312 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 77, + 505, + 144 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 77, + 506, + 92 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 506, + 92 + ], + "score": 1.0, + "content": "Table 2: 10-shot video captioning results. ♠ indicates concurrent work. The reported Flamingo [2]", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 89, + 506, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 89, + 506, + 102 + ], + "score": 1.0, + "content": "results are using 16 shots. #VideoPT represents the number of videos used for pre-training. B-4, R-L,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 99, + 506, + 112 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 116, + 110 + ], + "score": 0.3, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 99, + 120, + 112 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 120, + 100, + 128, + 110 + ], + "score": 0.53, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 99, + 372, + 112 + ], + "score": 1.0, + "content": "represents BLEU-4, ROUGE-L, METEOR and CIDEr. Avg", + "type": "text" + }, + { + "bbox": [ + 373, + 100, + 382, + 110 + ], + "score": 0.29, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 99, + 506, + 112 + ], + "score": 1.0, + "content": "represents the average CIDEr", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "score across all available benchmarks. ASR indicates whether the model has access to the ASR", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 120, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 145, + 135 + ], + "score": 1.0, + "content": "subtitles.", + "type": "text" + }, + { + "bbox": [ + 145, + 122, + 169, + 132 + ], + "score": 0.34, + "content": "B L I P", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 120, + 186, + 135 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 186, + 122, + 221, + 134 + ], + "score": 0.91, + "content": "B L I P _ { c a p }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 120, + 506, + 135 + ], + "score": 1.0, + "content": "use the pretrained checkpoint and the finetuned checkpoint on COCO", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 133, + 352, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 133, + 352, + 145 + ], + "score": 1.0, + "content": "captioning. All results are averaged over three random seeds.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "table_body", + "bbox": [ + 106, + 150, + 505, + 312 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 150, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 505, + 312 + ], + "score": 0.945, + "html": "
Method#VideOpr ASRB-4R-LMMSR-VTTCaptionYouCook2 Caption CVaTex CaptionAvg C
CB-4R-LMB-4 R-L MC
Few-shot
UniVL1.2MNo2.1 22.5 9.53.63.325.3 11.634.11.715.78.02.113.3
BLIP0No27.7 43.0 23.0 39.50.79.0 3.411.513.539.515.420.723.9
BLIPcap0No21.648.022.7 30.23.78.6 3.89.420.741.517.428.922.8
VidIL(ours)0No26.0 51.7 24.7 36.32.622.9 9.527.022.243.620.036.733.3
UniVL1.2MYes-=-4.326.4 12.248.62.717.710.23.426.0
VidIL(ours)0Yes10.7 35.919.4 111.623.2 44.2 20.6 38.975.3
Flamingo-3B(16)27MNo73.257.1
Flamingo-80B(16) 27MNo84.262.8=
Fine-tuning
UniVL1.2MNo42.0 61.0 29.0 50.1[11.2 40.1 17.6 127.0|22.8 38.6 22.3 33.470.2
UniVL1.2MYes=16.6 45.7 21.6176.823.7 39.3 22.7 35.6106.2
", + "type": "table", + "image_path": "bdc0755cadf49e276175a4fa571481c0553518341285949af409281797b235b2.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 106, + 150, + 505, + 204.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 106, + 204.0, + 505, + 258.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 106, + 258.0, + 505, + 312.0 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 4.75 + }, + { + "type": "text", + "bbox": [ + 106, + 333, + 505, + 377 + ], + "lines": [ + { + "bbox": [ + 105, + 333, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 345 + ], + "score": 1.0, + "content": "As discussed in Section 1, video captions describe the content in various semantic levels. The N-gram", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 343, + 507, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 507, + 357 + ], + "score": 1.0, + "content": "based metric may not fairly reflect the models’ performance in capturing the video-caption alignment.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "We further verify this hypothesis in Section 4.5. Thus, in addition to automatic metrics, we include", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 366, + 428, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 428, + 378 + ], + "score": 1.0, + "content": "qualitative examples illustrated in Figure 4. More examples are in Appendix ??.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 333, + 507, + 378 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 382, + 505, + 448 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 506, + 395 + ], + "score": 1.0, + "content": "Additionally, for most existing methods and also concurrent work, e.g., Flamingo [2], adding a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 393, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 505, + 405 + ], + "score": 1.0, + "content": "new modality often requires a dedicated model redesign or retraining. However, the nature of our", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 404, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 505, + 416 + ], + "score": 1.0, + "content": "framework, where we use a unified textual representation for each level, makes it highly flexible for", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 415, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 505, + 426 + ], + "score": 1.0, + "content": "incorporating new modalities. As shown in row 6 in Table, our model can effectively utilize extra", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 425, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 506, + 438 + ], + "score": 1.0, + "content": "information from ASR to obtain significantly better few-shot performance on certain datasets such as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 435, + 153, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 153, + 448 + ], + "score": 1.0, + "content": "YouCook2.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 381, + 506, + 448 + ] + }, + { + "type": "image", + "bbox": [ + 106, + 456, + 504, + 599 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 456, + 504, + 599 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 456, + 504, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 504, + 599 + ], + "score": 0.577, + "type": "image", + "image_path": "42488e5e146be037d7c57abcc48a78f8af1cf75e6517bc50d3e1b32bfd715b6a.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 106, + 456, + 504, + 503.6666666666667 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 106, + 503.6666666666667, + 504, + 551.3333333333334 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 106, + 551.3333333333334, + 504, + 599.0 + ], + "spans": [], + "index": 21 + } + ] + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 609, + 506, + 654 + ], + "lines": [ + { + "bbox": [ + 105, + 609, + 507, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 507, + 622 + ], + "score": 1.0, + "content": "Figure 4: Qualitative examples on video captioning. Grey boxes contain part of the video repre-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 619, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 633 + ], + "score": 1.0, + "content": "sentation from our model. Blue boxes contain caption generation from different models. Green", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 631, + 506, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 506, + 643 + ], + "score": 1.0, + "content": "boxes contain ground truth annotations. Bold green text highlights the correct information that is not", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 642, + 487, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 487, + 655 + ], + "score": 1.0, + "content": "captured in baseline outputs which can be reasoned from our visual tokens and frame captions.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 609, + 507, + 655 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 668, + 285, + 681 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 287, + 683 + ], + "score": 1.0, + "content": "4.3 Few-shot Video Question Answering", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 689, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 422, + 702 + ], + "score": 1.0, + "content": "We compare the test accuracy of our approach with few-shot pretrained BLIP,", + "type": "text" + }, + { + "bbox": [ + 423, + 689, + 465, + 701 + ], + "score": 0.87, + "content": "\\mathsf { B L I P } _ { V Q A }", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "[26], and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "concurrent work Flamingo [2] on two video question answering benchmarks, MSR-VTT_QA and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 159, + 723 + ], + "score": 1.0, + "content": "MSVD_QA.", + "type": "text" + }, + { + "bbox": [ + 160, + 711, + 201, + 723 + ], + "score": 0.76, + "content": "\\mathsf { B L I P } _ { V Q A }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "represents finetuned BLIP on VQA [4] dataset, which is the previous SOTA", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 444, + 86 + ], + "score": 1.0, + "content": "on zero/few-shot video question answering. In order to have fairer comparison with", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 445, + 73, + 487, + 85 + ], + "score": 0.89, + "content": "\\mathsf { B L I P } _ { V Q A }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 487, + 72, + 506, + 86 + ], + "score": 1.0, + "content": ", we", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 506, + 97 + ], + "score": 1.0, + "content": "reduce the shot number to 5 and report the average accuracy on three sets of randomly selected", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 93, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 104, + 93, + 506, + 108 + ], + "score": 1.0, + "content": "5-shot examples. As shown in Table 3, our method outperforms previous SOTA by a large margin.", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 106, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 506, + 118 + ], + "score": 1.0, + "content": "Comparing with concurrent work Flamingo, which is post-pretrained on a large number of video-text", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 114, + 507, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 507, + 130 + ], + "score": 1.0, + "content": "data, our model is training-free and did not observe any video data. However, with only image-", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "score": 1.0, + "content": "language and language-only knowledge, our 5-shot model is able to outperform 8-shot Flamingo-3B", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 138, + 347, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 347, + 150 + ], + "score": 1.0, + "content": "and achieve on-par performance with 4-shot Flamingo-80B.", + "type": "text", + "cross_page": true + } + ], + "index": 6 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 687, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 506, + 150 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 444, + 86 + ], + "score": 1.0, + "content": "on zero/few-shot video question answering. In order to have fairer comparison with", + "type": "text" + }, + { + "bbox": [ + 445, + 73, + 487, + 85 + ], + "score": 0.89, + "content": "\\mathsf { B L I P } _ { V Q A }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 72, + 506, + 86 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 506, + 97 + ], + "score": 1.0, + "content": "reduce the shot number to 5 and report the average accuracy on three sets of randomly selected", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 93, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 104, + 93, + 506, + 108 + ], + "score": 1.0, + "content": "5-shot examples. As shown in Table 3, our method outperforms previous SOTA by a large margin.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 106, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 506, + 118 + ], + "score": 1.0, + "content": "Comparing with concurrent work Flamingo, which is post-pretrained on a large number of video-text", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 114, + 507, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 507, + 130 + ], + "score": 1.0, + "content": "data, our model is training-free and did not observe any video data. However, with only image-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "score": 1.0, + "content": "language and language-only knowledge, our 5-shot model is able to outperform 8-shot Flamingo-3B", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 138, + 347, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 347, + 150 + ], + "score": 1.0, + "content": "and achieve on-par performance with 4-shot Flamingo-80B.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 108, + 189, + 376, + 306 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 160, + 378, + 183 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 159, + 380, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 214, + 173 + ], + "score": 1.0, + "content": "Table 3: Video QA results.", + "type": "text" + }, + { + "bbox": [ + 214, + 161, + 255, + 173 + ], + "score": 0.9, + "content": "\\mathsf { B L I P } _ { V Q A }", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 159, + 380, + 173 + ], + "score": 1.0, + "content": "is finetuned on VQA [4]. ♠ in-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 169, + 379, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 379, + 186 + ], + "score": 1.0, + "content": "dicates concurrent work. PT, FT indicates pretraining and finetuning.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.0 + }, + { + "type": "table_body", + "bbox": [ + 108, + 189, + 376, + 306 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 189, + 376, + 306 + ], + "spans": [ + { + "bbox": [ + 108, + 189, + 376, + 306 + ], + "score": 0.98, + "html": "
Method#videoPT#videoFTMSR-VTTMSVD
BLIP0O-shot0.550.45
BLIP05-shot0.840.53
BLIPvQA [26]0O-shot19.235.2
VidIL(ours)05-shot21.239.1
Flamingo-3B [2]27M4-shot14.933.0
Flamingo-3B_[2]27M8-shot19.637.0
Flamingo-80B [2]27M4-shot23.941.7
Flamingo-80B [2]27M8-shot27.645.5
ALPRO [25]2Mfull-shot42.145.9
", + "type": "table", + "image_path": "80c20213acc3c9f54437b9d3f7c1181c8608963e7d4b561375c9c1f29b5cc5cb.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 108, + 189, + 376, + 228.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 228.0, + 376, + 267.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 108, + 267.0, + 376, + 306.0 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 10.5 + }, + { + "type": "table", + "bbox": [ + 383, + 190, + 506, + 258 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 384, + 161, + 504, + 183 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 383, + 160, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 383, + 160, + 472, + 174 + ], + "score": 1.0, + "content": "Table 4: Accuracy", + "type": "text" + }, + { + "bbox": [ + 472, + 162, + 488, + 173 + ], + "score": 0.75, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 160, + 505, + 174 + ], + "score": 1.0, + "content": "on", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 383, + 171, + 473, + 184 + ], + "spans": [ + { + "bbox": [ + 383, + 171, + 473, + 184 + ], + "score": 1.0, + "content": "VLEP hidden test set.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.0 + }, + { + "type": "table_body", + "bbox": [ + 383, + 190, + 506, + 258 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 383, + 190, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 383, + 190, + 506, + 258 + ], + "score": 0.962, + "html": "
Method#videoFTAcc
VLEP [23] MERLOT[68]20142 2014267.5 68.4
VidIL(ours)10-shot72.0
Human-90.5
", + "type": "table", + "image_path": "73514530509fe457864a28260c85bbc7f3012661dc1bd0d43e115e6c87eefb5c.jpg" + } + ] + } + ], + "index": 13.0, + "virtual_lines": [ + { + "bbox": [ + 383, + 190, + 506, + 224.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 383, + 224.0, + 506, + 258.0 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 11.0 + }, + { + "type": "title", + "bbox": [ + 108, + 320, + 315, + 332 + ], + "lines": [ + { + "bbox": [ + 105, + 319, + 316, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 316, + 334 + ], + "score": 1.0, + "content": "4.4 Few-shot Video-Language Event Prediction", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 340, + 505, + 493 + ], + "lines": [ + { + "bbox": [ + 105, + 339, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 506, + 353 + ], + "score": 1.0, + "content": "In this section, we show that our model not only can answer questions about the video visual features", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 351, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 506, + 363 + ], + "score": 1.0, + "content": "but also answering \"What is more likely to happen next?\". Given a video with associated subtitle", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "transcript as premise, the video-language event prediction (VLEP) task is to predict the most likely", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 372, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 386 + ], + "score": 1.0, + "content": "future event. The original VLEP [23] paper formulates the problem as a binary classification problem", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 383, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 505, + 396 + ], + "score": 1.0, + "content": "where the model will be chosen from two possible future event candidates. Instead, we formulate this", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 395, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 505, + 408 + ], + "score": 1.0, + "content": "problem as another video-to-text generation problem to fit into our framework. Figure 5 depicts an", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 406, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 506, + 419 + ], + "score": 1.0, + "content": "example with the same format as in Figure 2. Similar to the evaluation setting in QA, the generated", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 416, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 506, + 429 + ], + "score": 1.0, + "content": "free-form text will first be mapped to one of the two candidate answers using SentenceBert [47], and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 426, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 506, + 441 + ], + "score": 1.0, + "content": "then calculate the accuracy. In Table 4, we report accuracy on the hidden test set of VLEP [23]. To our", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 438, + 507, + 452 + ], + "spans": [ + { + "bbox": [ + 104, + 438, + 507, + 452 + ], + "score": 1.0, + "content": "surprise, our 10-shot model outperforms state-of-the-art fully-supervised baseline, i.e., MERLOT [68],", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 449, + 507, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 179, + 462 + ], + "score": 1.0, + "content": "by a large margin", + "type": "text" + }, + { + "bbox": [ + 180, + 449, + 209, + 460 + ], + "score": 0.86, + "content": "( \\sim 4 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 449, + 507, + 462 + ], + "score": 1.0, + "content": ". This shows that our model has strong few-shot ability not only on video-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 460, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 473 + ], + "score": 1.0, + "content": "language understanding but also on prediction. Since event prediction tasks rely heavily on temporal", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "score": 1.0, + "content": "ordering, we show that with the proposed temporal-aware prompting, language models can be guided", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 482, + 376, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 376, + 495 + ], + "score": 1.0, + "content": "to capture temporal dynamics between historical and future events.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 23.5 + }, + { + "type": "image", + "bbox": [ + 115, + 504, + 493, + 584 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 504, + 493, + 584 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 504, + 493, + 584 + ], + "spans": [ + { + "bbox": [ + 115, + 504, + 493, + 584 + ], + "score": 0.893, + "type": "image", + "image_path": "337717ccf8591fcb87cdc2bcb90b9982ea93bafdcb263be40f57ceb4066b57ea.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 115, + 504, + 493, + 530.6666666666666 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 115, + 530.6666666666666, + 493, + 557.3333333333333 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 115, + 557.3333333333333, + 493, + 583.9999999999999 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 244, + 587, + 367, + 599 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 243, + 586, + 367, + 600 + ], + "spans": [ + { + "bbox": [ + 243, + 586, + 367, + 600 + ], + "score": 1.0, + "content": "Figure 5: Prompt for VLEP task.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + } + ], + "index": 33.0 + }, + { + "type": "title", + "bbox": [ + 108, + 614, + 291, + 626 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 292, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 292, + 627 + ], + "score": 1.0, + "content": "4.5 Semi-supervised Text-Video Retrieval", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 634, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 634, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 505, + 647 + ], + "score": 1.0, + "content": "In addition to video-to-text generation tasks, we show that a broader range of video-language tasks", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "can benefit from our few-shot video captioner from a data perspective. Here, we consider a low-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 654, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 669 + ], + "score": 1.0, + "content": "budget semi-supervised setting where we only have a few labeled video-caption pairs and a large", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "amount of unlabeled videos. The idea is to leverage our video captioner to generate pseudo labels for", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 678, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 506, + 690 + ], + "score": 1.0, + "content": "training any given vision-language models. As a case study, we evaluate on two text-video retrieval", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "benchmarks, i.e., MSR-VTT and VaTeX. We use greedy decoding to generate pseudo caption for", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "each video in the training set. We then train an identical base model, i.e., BLIP, using different pseudo", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 373, + 723 + ], + "score": 1.0, + "content": "labeled data as well as ground truth annotations. We report Recall", + "type": "text" + }, + { + "bbox": [ + 374, + 711, + 383, + 721 + ], + "score": 0.71, + "content": "@", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "1 and 5 for both video-to-text", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 39.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 506, + 150 + ], + "lines": [], + "index": 3, + "bbox_fs": [ + 104, + 72, + 507, + 150 + ], + "lines_deleted": true + }, + { + "type": "table", + "bbox": [ + 108, + 189, + 376, + 306 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 160, + 378, + 183 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 159, + 380, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 214, + 173 + ], + "score": 1.0, + "content": "Table 3: Video QA results.", + "type": "text" + }, + { + "bbox": [ + 214, + 161, + 255, + 173 + ], + "score": 0.9, + "content": "\\mathsf { B L I P } _ { V Q A }", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 159, + 380, + 173 + ], + "score": 1.0, + "content": "is finetuned on VQA [4]. ♠ in-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 169, + 379, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 379, + 186 + ], + "score": 1.0, + "content": "dicates concurrent work. PT, FT indicates pretraining and finetuning.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.0 + }, + { + "type": "table_body", + "bbox": [ + 108, + 189, + 376, + 306 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 189, + 376, + 306 + ], + "spans": [ + { + "bbox": [ + 108, + 189, + 376, + 306 + ], + "score": 0.98, + "html": "
Method#videoPT#videoFTMSR-VTTMSVD
BLIP0O-shot0.550.45
BLIP05-shot0.840.53
BLIPvQA [26]0O-shot19.235.2
VidIL(ours)05-shot21.239.1
Flamingo-3B [2]27M4-shot14.933.0
Flamingo-3B_[2]27M8-shot19.637.0
Flamingo-80B [2]27M4-shot23.941.7
Flamingo-80B [2]27M8-shot27.645.5
ALPRO [25]2Mfull-shot42.145.9
", + "type": "table", + "image_path": "80c20213acc3c9f54437b9d3f7c1181c8608963e7d4b561375c9c1f29b5cc5cb.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 108, + 189, + 376, + 228.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 228.0, + 376, + 267.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 108, + 267.0, + 376, + 306.0 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 10.5 + }, + { + "type": "table", + "bbox": [ + 383, + 190, + 506, + 258 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 384, + 161, + 504, + 183 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 383, + 160, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 383, + 160, + 472, + 174 + ], + "score": 1.0, + "content": "Table 4: Accuracy", + "type": "text" + }, + { + "bbox": [ + 472, + 162, + 488, + 173 + ], + "score": 0.75, + "content": "( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 160, + 505, + 174 + ], + "score": 1.0, + "content": "on", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 383, + 171, + 473, + 184 + ], + "spans": [ + { + "bbox": [ + 383, + 171, + 473, + 184 + ], + "score": 1.0, + "content": "VLEP hidden test set.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.0 + }, + { + "type": "table_body", + "bbox": [ + 383, + 190, + 506, + 258 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 383, + 190, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 383, + 190, + 506, + 258 + ], + "score": 0.962, + "html": "
Method#videoFTAcc
VLEP [23] MERLOT[68]20142 2014267.5 68.4
VidIL(ours)10-shot72.0
Human-90.5
", + "type": "table", + "image_path": "73514530509fe457864a28260c85bbc7f3012661dc1bd0d43e115e6c87eefb5c.jpg" + } + ] + } + ], + "index": 13.0, + "virtual_lines": [ + { + "bbox": [ + 383, + 190, + 506, + 224.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 383, + 224.0, + 506, + 258.0 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 11.0 + }, + { + "type": "title", + "bbox": [ + 108, + 320, + 315, + 332 + ], + "lines": [ + { + "bbox": [ + 105, + 319, + 316, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 316, + 334 + ], + "score": 1.0, + "content": "4.4 Few-shot Video-Language Event Prediction", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 340, + 505, + 493 + ], + "lines": [ + { + "bbox": [ + 105, + 339, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 506, + 353 + ], + "score": 1.0, + "content": "In this section, we show that our model not only can answer questions about the video visual features", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 351, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 506, + 363 + ], + "score": 1.0, + "content": "but also answering \"What is more likely to happen next?\". Given a video with associated subtitle", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "transcript as premise, the video-language event prediction (VLEP) task is to predict the most likely", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 372, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 386 + ], + "score": 1.0, + "content": "future event. The original VLEP [23] paper formulates the problem as a binary classification problem", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 383, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 505, + 396 + ], + "score": 1.0, + "content": "where the model will be chosen from two possible future event candidates. Instead, we formulate this", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 395, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 505, + 408 + ], + "score": 1.0, + "content": "problem as another video-to-text generation problem to fit into our framework. Figure 5 depicts an", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 406, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 506, + 419 + ], + "score": 1.0, + "content": "example with the same format as in Figure 2. Similar to the evaluation setting in QA, the generated", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 416, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 506, + 429 + ], + "score": 1.0, + "content": "free-form text will first be mapped to one of the two candidate answers using SentenceBert [47], and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 426, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 506, + 441 + ], + "score": 1.0, + "content": "then calculate the accuracy. In Table 4, we report accuracy on the hidden test set of VLEP [23]. To our", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 438, + 507, + 452 + ], + "spans": [ + { + "bbox": [ + 104, + 438, + 507, + 452 + ], + "score": 1.0, + "content": "surprise, our 10-shot model outperforms state-of-the-art fully-supervised baseline, i.e., MERLOT [68],", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 449, + 507, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 179, + 462 + ], + "score": 1.0, + "content": "by a large margin", + "type": "text" + }, + { + "bbox": [ + 180, + 449, + 209, + 460 + ], + "score": 0.86, + "content": "( \\sim 4 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 449, + 507, + 462 + ], + "score": 1.0, + "content": ". This shows that our model has strong few-shot ability not only on video-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 460, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 473 + ], + "score": 1.0, + "content": "language understanding but also on prediction. Since event prediction tasks rely heavily on temporal", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "score": 1.0, + "content": "ordering, we show that with the proposed temporal-aware prompting, language models can be guided", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 482, + 376, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 376, + 495 + ], + "score": 1.0, + "content": "to capture temporal dynamics between historical and future events.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 23.5, + "bbox_fs": [ + 104, + 339, + 507, + 495 + ] + }, + { + "type": "image", + "bbox": [ + 115, + 504, + 493, + 584 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 504, + 493, + 584 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 504, + 493, + 584 + ], + "spans": [ + { + "bbox": [ + 115, + 504, + 493, + 584 + ], + "score": 0.893, + "type": "image", + "image_path": "337717ccf8591fcb87cdc2bcb90b9982ea93bafdcb263be40f57ceb4066b57ea.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 115, + 504, + 493, + 530.6666666666666 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 115, + 530.6666666666666, + 493, + 557.3333333333333 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 115, + 557.3333333333333, + 493, + 583.9999999999999 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 244, + 587, + 367, + 599 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 243, + 586, + 367, + 600 + ], + "spans": [ + { + "bbox": [ + 243, + 586, + 367, + 600 + ], + "score": 1.0, + "content": "Figure 5: Prompt for VLEP task.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + } + ], + "index": 33.0 + }, + { + "type": "title", + "bbox": [ + 108, + 614, + 291, + 626 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 292, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 292, + 627 + ], + "score": 1.0, + "content": "4.5 Semi-supervised Text-Video Retrieval", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 634, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 634, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 505, + 647 + ], + "score": 1.0, + "content": "In addition to video-to-text generation tasks, we show that a broader range of video-language tasks", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "can benefit from our few-shot video captioner from a data perspective. Here, we consider a low-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 654, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 669 + ], + "score": 1.0, + "content": "budget semi-supervised setting where we only have a few labeled video-caption pairs and a large", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "amount of unlabeled videos. The idea is to leverage our video captioner to generate pseudo labels for", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 678, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 506, + 690 + ], + "score": 1.0, + "content": "training any given vision-language models. As a case study, we evaluate on two text-video retrieval", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "benchmarks, i.e., MSR-VTT and VaTeX. We use greedy decoding to generate pseudo caption for", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "each video in the training set. We then train an identical base model, i.e., BLIP, using different pseudo", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 373, + 723 + ], + "score": 1.0, + "content": "labeled data as well as ground truth annotations. We report Recall", + "type": "text" + }, + { + "bbox": [ + 374, + 711, + 383, + 721 + ], + "score": 0.71, + "content": "@", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "1 and 5 for both video-to-text", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 230, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 506, + 244 + ], + "score": 1.0, + "content": "and text-to-video retrieval. Table 5 shows that through training on our pseudo labels, we can achieve", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "significant improvements compared with zero-shot BLIP. We also show that the performance gain", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "score": 1.0, + "content": "is not simply a result of training on more data, since finetuning on the pseudo labels generated by", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 264, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 276 + ], + "score": 1.0, + "content": "other baselines (UniVL, BLIP) is less effective and can even hurt the performance. Furthermore, on", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 274, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 181, + 287 + ], + "score": 1.0, + "content": "MSR-VTT Recall", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 181, + 275, + 190, + 285 + ], + "score": 0.59, + "content": "@", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 191, + 274, + 506, + 287 + ], + "score": 1.0, + "content": "5 we can even achieve comparable performance against BLIP model finetuned", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 286, + 238, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 238, + 298 + ], + "score": 1.0, + "content": "on full ground truth annotations.", + "type": "text", + "cross_page": true + } + ], + "index": 11 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 634, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 110, + 505, + 213 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 70, + 505, + 104 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 68, + 505, + 83 + ], + "spans": [ + { + "bbox": [ + 104, + 68, + 407, + 83 + ], + "score": 1.0, + "content": "Table 5: Semi-supervised text-video retrieval with 10 labeled examples.", + "type": "text" + }, + { + "bbox": [ + 407, + 70, + 430, + 82 + ], + "score": 0.87, + "content": "\\mathrm { V _ { l a b e l } }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 68, + 443, + 83 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 443, + 70, + 473, + 82 + ], + "score": 0.88, + "content": "\\mathrm { \\Delta V _ { u n l a b e l } }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 68, + 505, + 83 + ], + "score": 1.0, + "content": "are the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 81, + 505, + 93 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 324, + 93 + ], + "score": 1.0, + "content": "number of labeled and unlabeled videos, respectively.", + "type": "text" + }, + { + "bbox": [ + 325, + 82, + 345, + 92 + ], + "score": 0.65, + "content": "t \\_ R I", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 81, + 363, + 93 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 363, + 82, + 379, + 93 + ], + "score": 0.79, + "content": "t \\_ R", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 81, + 490, + 93 + ], + "score": 1.0, + "content": "denote video-to-text Recall", + "type": "text" + }, + { + "bbox": [ + 490, + 82, + 505, + 91 + ], + "score": 0.69, + "content": "@ 1", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 92, + 348, + 104 + ], + "spans": [ + { + "bbox": [ + 106, + 92, + 133, + 103 + ], + "score": 1.0, + "content": "and 5.", + "type": "text" + }, + { + "bbox": [ + 134, + 93, + 155, + 104 + ], + "score": 0.53, + "content": "\\nu \\_ R I", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 92, + 173, + 103 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 174, + 93, + 196, + 104 + ], + "score": 0.73, + "content": "\\nu \\_ R 5", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 92, + 306, + 103 + ], + "score": 1.0, + "content": "denote text-to-video Recall", + "type": "text" + }, + { + "bbox": [ + 306, + 93, + 321, + 102 + ], + "score": 0.73, + "content": "@ 1", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 92, + 348, + 103 + ], + "score": 1.0, + "content": "and 5.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 106, + 110, + 505, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 110, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 505, + 213 + ], + "score": 0.984, + "html": "
ModelPseudo LabelMSR-VTTRetrievalVaTexRetrieval
Vlabel/Vunlabel t_R1t_R5v_R1v_R5Vlabel/Vunlabel t_R1t_R5v_R1v_R5
BLIP33.257.240.562.828.253.434.058.6
BLIPUniVL10 /701033.157.333.657.710 /2268525.547.726.149.1
BLIPBLIP10/701035.660.839.860.410/2268526.350.529.353.6
BLIPBLIPcap10/701035.358.039.163.310/2268523.946.827.549.7
BLIPVidIL(ours)10/701039.664.540.865.210 /2268533.359.133.759.5
BLIPGround Truth7010/043.666.243.167.222685/040.166.440.166.6
ALPRO [25]Ground Truth140200/032.060.633.960.7----
DRL [56]Ground Truth180000/054.177.452.978.5-----
", + "type": "table", + "image_path": "3d91cbc690fa88f5cb9c62c6176d417056eb1bb8034ac54fa9a50416e7a1be80.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 106, + 110, + 505, + 144.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 106, + 144.33333333333334, + 505, + 178.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 106, + 178.66666666666669, + 505, + 213.00000000000003 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 231, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 230, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 506, + 244 + ], + "score": 1.0, + "content": "and text-to-video retrieval. Table 5 shows that through training on our pseudo labels, we can achieve", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "significant improvements compared with zero-shot BLIP. We also show that the performance gain", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "score": 1.0, + "content": "is not simply a result of training on more data, since finetuning on the pseudo labels generated by", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 264, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 276 + ], + "score": 1.0, + "content": "other baselines (UniVL, BLIP) is less effective and can even hurt the performance. Furthermore, on", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 274, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 181, + 287 + ], + "score": 1.0, + "content": "MSR-VTT Recall", + "type": "text" + }, + { + "bbox": [ + 181, + 275, + 190, + 285 + ], + "score": 0.59, + "content": "@", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 274, + 506, + 287 + ], + "score": 1.0, + "content": "5 we can even achieve comparable performance against BLIP model finetuned", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 286, + 238, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 238, + 298 + ], + "score": 1.0, + "content": "on full ground truth annotations.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 302, + 506, + 412 + ], + "lines": [ + { + "bbox": [ + 105, + 301, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 506, + 315 + ], + "score": 1.0, + "content": "Another interesting observation is that, compared with the video captioning results in Table 2, we", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 313, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 506, + 325 + ], + "score": 1.0, + "content": "found that the gain of our model over baselines on text-video retrieval is more visible than on", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 324, + 506, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 506, + 337 + ], + "score": 1.0, + "content": "captioning. A key factor in performing well on text-video retrieval tasks is to learn a good video-text", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 334, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 348 + ], + "score": 1.0, + "content": "multi-modal alignment. This result shows that our pseudo labels capture richer video-text alignment", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 345, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 359 + ], + "score": 1.0, + "content": "that can benefit the retrieval-style downstream task. The N-gram based generation metrics, e.g.,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "score": 1.0, + "content": "BLEU, may not be able to fully reflect the alignment information, due to the variety of semantic", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "score": 1.0, + "content": "levels in video captions. Furthermore, from a data perspective, our video captioner can be viewed", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 379, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 391 + ], + "score": 1.0, + "content": "as a data augmentation tool which is capable of generating or augmenting any open-domain video-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 390, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 403 + ], + "score": 1.0, + "content": "language pretraining datasets with minimal human effort. As a result, we can potentially improve", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 400, + 464, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 464, + 414 + ], + "score": 1.0, + "content": "video-language pretraining by constructing a cleaner and more diverse video-text corpus.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 16.5 + }, + { + "type": "table", + "bbox": [ + 109, + 456, + 316, + 554 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 427, + 319, + 449 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 426, + 320, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 320, + 439 + ], + "score": 1.0, + "content": "Table 6: Impact of visual tokens and temporal dimen-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 438, + 129, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 129, + 450 + ], + "score": 1.0, + "content": "sion.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "table_body", + "bbox": [ + 109, + 456, + 316, + 554 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 109, + 456, + 316, + 554 + ], + "spans": [ + { + "bbox": [ + 109, + 456, + 316, + 554 + ], + "score": 0.979, + "html": "
Video RepresentationAvg↑ Std↓
Visual TokenFrame Frame+Object39.6 40.33.7 2.9
Frame+Object+Event39.92.8
Frame+Object+Attibute Frame+Object+Event+Attribute40.9 40.82.9 2.4
TemporalReduce to one frame Reverse temporal order38.5 40.72.4 1.7
", + "type": "table", + "image_path": "1216fec2ba1e8eca8e72af10e7c75e06a1a18c33372a4215061d9f781eb347cd.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 109, + 456, + 316, + 470.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 109, + 470.0, + 316, + 484.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 109, + 484.0, + 316, + 498.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 109, + 498.0, + 316, + 512.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 109, + 512.0, + 316, + 526.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 109, + 526.0, + 316, + 540.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 109, + 540.0, + 316, + 554.0 + ], + "spans": [], + "index": 38 + } + ] + } + ], + "index": 27.25 + }, + { + "type": "table", + "bbox": [ + 323, + 481, + 505, + 553 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 322, + 434, + 504, + 479 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 323, + 435, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 323, + 435, + 505, + 446 + ], + "score": 1.0, + "content": "Table 7: Impact of shot selection. #ICE in-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 323, + 445, + 504, + 457 + ], + "spans": [ + { + "bbox": [ + 323, + 445, + 504, + 457 + ], + "score": 1.0, + "content": "dicates the number of in-context examples in", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 323, + 456, + 504, + 469 + ], + "spans": [ + { + "bbox": [ + 323, + 456, + 504, + 469 + ], + "score": 1.0, + "content": "the prompt. Details of in-context example", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 323, + 468, + 444, + 480 + ], + "spans": [ + { + "bbox": [ + 323, + 468, + 444, + 480 + ], + "score": 1.0, + "content": "selection are in the Appendix.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.0 + }, + { + "type": "table_body", + "bbox": [ + 323, + 481, + 505, + 553 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 323, + 481, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 323, + 481, + 505, + 553 + ], + "score": 0.978, + "html": "
#shotw/o selectionw/ selection
#ICEAvg↑Std#ICEAvg↑Std↓
5538.42.1540.41.2
101041.33.6540.82.4
202042.63.3542.22.0
303040.02.9541.11.9
", + "type": "table", + "image_path": "b6823d5dd736e75f8f7e9c3d796ad639cd35817997264d469a99028f8ff5a52a.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 323, + 481, + 505, + 495.4 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 323, + 495.4, + 505, + 509.79999999999995 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 323, + 509.79999999999995, + 505, + 524.1999999999999 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 323, + 524.1999999999999, + 505, + 538.5999999999999 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 323, + 538.5999999999999, + 505, + 552.9999999999999 + ], + "spans": [], + "index": 39 + } + ] + } + ], + "index": 30.5 + }, + { + "type": "title", + "bbox": [ + 107, + 584, + 201, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 583, + 203, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 203, + 597 + ], + "score": 1.0, + "content": "4.6 Ablation Studies", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 106, + 606, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 106, + 607, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 506, + 619 + ], + "score": 1.0, + "content": "We perform comprehensive ablation studies on our few-shot prompt including the impact of different", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "video representation, number of shots and in-context selection. All the ablation results are evaluated", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "on MSVD_QA validation set, and we report the mean and standard deviation of each setting on three", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 640, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 506, + 651 + ], + "score": 1.0, + "content": "sets of randomly sampled shots. For the cases with in-context example selection, we further select", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 651, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 505, + 662 + ], + "score": 1.0, + "content": "5 examples as in-context examples from the sampled shots, while for the cases without in-context", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 662, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 505, + 674 + ], + "score": 1.0, + "content": "selection, all shots will be feed into the prompt. In Table 6, we show adding visual tokens consistently", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 673, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 505, + 685 + ], + "score": 1.0, + "content": "improves not only the model accuracy but also the model variance. A lower standard deviation", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 682, + 371, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 371, + 696 + ], + "score": 1.0, + "content": "indicates that the model is less sensitive to the few-shot sampling.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 503, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "To further demonstrate the impact of the additional temporal dimension of videos, we perform two", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "ablations on the \"Frame+Object+Event+Attribute\" setting. First, we reduce the number of", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 110, + 505, + 213 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 70, + 505, + 104 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 68, + 505, + 83 + ], + "spans": [ + { + "bbox": [ + 104, + 68, + 407, + 83 + ], + "score": 1.0, + "content": "Table 5: Semi-supervised text-video retrieval with 10 labeled examples.", + "type": "text" + }, + { + "bbox": [ + 407, + 70, + 430, + 82 + ], + "score": 0.87, + "content": "\\mathrm { V _ { l a b e l } }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 68, + 443, + 83 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 443, + 70, + 473, + 82 + ], + "score": 0.88, + "content": "\\mathrm { \\Delta V _ { u n l a b e l } }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 68, + 505, + 83 + ], + "score": 1.0, + "content": "are the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 81, + 505, + 93 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 324, + 93 + ], + "score": 1.0, + "content": "number of labeled and unlabeled videos, respectively.", + "type": "text" + }, + { + "bbox": [ + 325, + 82, + 345, + 92 + ], + "score": 0.65, + "content": "t \\_ R I", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 81, + 363, + 93 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 363, + 82, + 379, + 93 + ], + "score": 0.79, + "content": "t \\_ R", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 81, + 490, + 93 + ], + "score": 1.0, + "content": "denote video-to-text Recall", + "type": "text" + }, + { + "bbox": [ + 490, + 82, + 505, + 91 + ], + "score": 0.69, + "content": "@ 1", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 92, + 348, + 104 + ], + "spans": [ + { + "bbox": [ + 106, + 92, + 133, + 103 + ], + "score": 1.0, + "content": "and 5.", + "type": "text" + }, + { + "bbox": [ + 134, + 93, + 155, + 104 + ], + "score": 0.53, + "content": "\\nu \\_ R I", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 92, + 173, + 103 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 174, + 93, + 196, + 104 + ], + "score": 0.73, + "content": "\\nu \\_ R 5", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 92, + 306, + 103 + ], + "score": 1.0, + "content": "denote text-to-video Recall", + "type": "text" + }, + { + "bbox": [ + 306, + 93, + 321, + 102 + ], + "score": 0.73, + "content": "@ 1", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 92, + 348, + 103 + ], + "score": 1.0, + "content": "and 5.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 106, + 110, + 505, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 110, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 505, + 213 + ], + "score": 0.984, + "html": "
ModelPseudo LabelMSR-VTTRetrievalVaTexRetrieval
Vlabel/Vunlabel t_R1t_R5v_R1v_R5Vlabel/Vunlabel t_R1t_R5v_R1v_R5
BLIP33.257.240.562.828.253.434.058.6
BLIPUniVL10 /701033.157.333.657.710 /2268525.547.726.149.1
BLIPBLIP10/701035.660.839.860.410/2268526.350.529.353.6
BLIPBLIPcap10/701035.358.039.163.310/2268523.946.827.549.7
BLIPVidIL(ours)10/701039.664.540.865.210 /2268533.359.133.759.5
BLIPGround Truth7010/043.666.243.167.222685/040.166.440.166.6
ALPRO [25]Ground Truth140200/032.060.633.960.7----
DRL [56]Ground Truth180000/054.177.452.978.5-----
", + "type": "table", + "image_path": "3d91cbc690fa88f5cb9c62c6176d417056eb1bb8034ac54fa9a50416e7a1be80.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 106, + 110, + 505, + 144.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 106, + 144.33333333333334, + 505, + 178.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 106, + 178.66666666666669, + 505, + 213.00000000000003 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 231, + 505, + 297 + ], + "lines": [], + "index": 8.5, + "bbox_fs": [ + 105, + 230, + 506, + 298 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 302, + 506, + 412 + ], + "lines": [ + { + "bbox": [ + 105, + 301, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 506, + 315 + ], + "score": 1.0, + "content": "Another interesting observation is that, compared with the video captioning results in Table 2, we", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 313, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 506, + 325 + ], + "score": 1.0, + "content": "found that the gain of our model over baselines on text-video retrieval is more visible than on", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 324, + 506, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 506, + 337 + ], + "score": 1.0, + "content": "captioning. A key factor in performing well on text-video retrieval tasks is to learn a good video-text", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 334, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 348 + ], + "score": 1.0, + "content": "multi-modal alignment. This result shows that our pseudo labels capture richer video-text alignment", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 345, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 359 + ], + "score": 1.0, + "content": "that can benefit the retrieval-style downstream task. The N-gram based generation metrics, e.g.,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 370 + ], + "score": 1.0, + "content": "BLEU, may not be able to fully reflect the alignment information, due to the variety of semantic", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "score": 1.0, + "content": "levels in video captions. Furthermore, from a data perspective, our video captioner can be viewed", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 379, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 391 + ], + "score": 1.0, + "content": "as a data augmentation tool which is capable of generating or augmenting any open-domain video-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 390, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 403 + ], + "score": 1.0, + "content": "language pretraining datasets with minimal human effort. As a result, we can potentially improve", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 400, + 464, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 464, + 414 + ], + "score": 1.0, + "content": "video-language pretraining by constructing a cleaner and more diverse video-text corpus.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 301, + 506, + 414 + ] + }, + { + "type": "table", + "bbox": [ + 109, + 456, + 316, + 554 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 427, + 319, + 449 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 426, + 320, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 320, + 439 + ], + "score": 1.0, + "content": "Table 6: Impact of visual tokens and temporal dimen-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 438, + 129, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 129, + 450 + ], + "score": 1.0, + "content": "sion.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "table_body", + "bbox": [ + 109, + 456, + 316, + 554 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 109, + 456, + 316, + 554 + ], + "spans": [ + { + "bbox": [ + 109, + 456, + 316, + 554 + ], + "score": 0.979, + "html": "
Video RepresentationAvg↑ Std↓
Visual TokenFrame Frame+Object39.6 40.33.7 2.9
Frame+Object+Event39.92.8
Frame+Object+Attibute Frame+Object+Event+Attribute40.9 40.82.9 2.4
TemporalReduce to one frame Reverse temporal order38.5 40.72.4 1.7
", + "type": "table", + "image_path": "1216fec2ba1e8eca8e72af10e7c75e06a1a18c33372a4215061d9f781eb347cd.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 109, + 456, + 316, + 470.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 109, + 470.0, + 316, + 484.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 109, + 484.0, + 316, + 498.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 109, + 498.0, + 316, + 512.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 109, + 512.0, + 316, + 526.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 109, + 526.0, + 316, + 540.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 109, + 540.0, + 316, + 554.0 + ], + "spans": [], + "index": 38 + } + ] + } + ], + "index": 27.25 + }, + { + "type": "table", + "bbox": [ + 323, + 481, + 505, + 553 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 322, + 434, + 504, + 479 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 323, + 435, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 323, + 435, + 505, + 446 + ], + "score": 1.0, + "content": "Table 7: Impact of shot selection. #ICE in-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 323, + 445, + 504, + 457 + ], + "spans": [ + { + "bbox": [ + 323, + 445, + 504, + 457 + ], + "score": 1.0, + "content": "dicates the number of in-context examples in", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 323, + 456, + 504, + 469 + ], + "spans": [ + { + "bbox": [ + 323, + 456, + 504, + 469 + ], + "score": 1.0, + "content": "the prompt. Details of in-context example", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 323, + 468, + 444, + 480 + ], + "spans": [ + { + "bbox": [ + 323, + 468, + 444, + 480 + ], + "score": 1.0, + "content": "selection are in the Appendix.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.0 + }, + { + "type": "table_body", + "bbox": [ + 323, + 481, + 505, + 553 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 323, + 481, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 323, + 481, + 505, + 553 + ], + "score": 0.978, + "html": "
#shotw/o selectionw/ selection
#ICEAvg↑Std#ICEAvg↑Std↓
5538.42.1540.41.2
101041.33.6540.82.4
202042.63.3542.22.0
303040.02.9541.11.9
", + "type": "table", + "image_path": "b6823d5dd736e75f8f7e9c3d796ad639cd35817997264d469a99028f8ff5a52a.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 323, + 481, + 505, + 495.4 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 323, + 495.4, + 505, + 509.79999999999995 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 323, + 509.79999999999995, + 505, + 524.1999999999999 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 323, + 524.1999999999999, + 505, + 538.5999999999999 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 323, + 538.5999999999999, + 505, + 552.9999999999999 + ], + "spans": [], + "index": 39 + } + ] + } + ], + "index": 30.5 + }, + { + "type": "title", + "bbox": [ + 107, + 584, + 201, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 583, + 203, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 203, + 597 + ], + "score": 1.0, + "content": "4.6 Ablation Studies", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 106, + 606, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 106, + 607, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 506, + 619 + ], + "score": 1.0, + "content": "We perform comprehensive ablation studies on our few-shot prompt including the impact of different", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "video representation, number of shots and in-context selection. All the ablation results are evaluated", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "on MSVD_QA validation set, and we report the mean and standard deviation of each setting on three", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 640, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 506, + 651 + ], + "score": 1.0, + "content": "sets of randomly sampled shots. For the cases with in-context example selection, we further select", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 651, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 505, + 662 + ], + "score": 1.0, + "content": "5 examples as in-context examples from the sampled shots, while for the cases without in-context", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 662, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 505, + 674 + ], + "score": 1.0, + "content": "selection, all shots will be feed into the prompt. In Table 6, we show adding visual tokens consistently", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 673, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 505, + 685 + ], + "score": 1.0, + "content": "improves not only the model accuracy but also the model variance. A lower standard deviation", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 682, + 371, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 371, + 696 + ], + "score": 1.0, + "content": "indicates that the model is less sensitive to the few-shot sampling.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 607, + 506, + 696 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 700, + 503, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "To further demonstrate the impact of the additional temporal dimension of videos, we perform two", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "ablations on the \"Frame+Object+Event+Attribute\" setting. First, we reduce the number of", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "score": 1.0, + "content": "frame captions and visual tokens to be one9 for each video. We found that the performance drops", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 84, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 84, + 505, + 96 + ], + "score": 1.0, + "content": "significantly compared with using the default four frames, which indicates the model’s ability to", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 95, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 505, + 107 + ], + "score": 1.0, + "content": "incorporate information from multiple timestamps. Further, we found that fine-grained temporal", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 106, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 505, + 117 + ], + "score": 1.0, + "content": "modeling is rarely required for performing well on current video-language benchmarks. As shown in", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 116, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 505, + 129 + ], + "score": 1.0, + "content": "the ablation result where we reverse the order of all visual tokens and frame captions, the performance", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 126, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 141 + ], + "score": 1.0, + "content": "decreased only marginally, which indicates that current benchmarks may not be sufficient in reflecting", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 137, + 278, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 278, + 151 + ], + "score": 1.0, + "content": "the benefits from better temporal ordering.", + "type": "text", + "cross_page": true + } + ], + "index": 6 + } + ], + "index": 49.5, + "bbox_fs": [ + 106, + 699, + 505, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "score": 1.0, + "content": "frame captions and visual tokens to be one9 for each video. We found that the performance drops", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 84, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 84, + 505, + 96 + ], + "score": 1.0, + "content": "significantly compared with using the default four frames, which indicates the model’s ability to", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 95, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 505, + 107 + ], + "score": 1.0, + "content": "incorporate information from multiple timestamps. Further, we found that fine-grained temporal", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 106, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 505, + 117 + ], + "score": 1.0, + "content": "modeling is rarely required for performing well on current video-language benchmarks. As shown in", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 116, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 505, + 129 + ], + "score": 1.0, + "content": "the ablation result where we reverse the order of all visual tokens and frame captions, the performance", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 126, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 141 + ], + "score": 1.0, + "content": "decreased only marginally, which indicates that current benchmarks may not be sufficient in reflecting", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 137, + 278, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 278, + 151 + ], + "score": 1.0, + "content": "the benefits from better temporal ordering.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 253 + ], + "lines": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "In Table 7, we first show that, with the same context length, namely, 5 in-context examples, in-context", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 165, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 177 + ], + "score": 1.0, + "content": "example selection significantly increases the performance as well as the robustness. At 10-shot, and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 177, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 506, + 190 + ], + "score": 1.0, + "content": "20-shot, directly fitting more shots into the prompt results in better performance. In-context selection", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 188, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 188, + 506, + 199 + ], + "score": 1.0, + "content": "achieves slightly lower performance but with significantly better efficiency due to shorter context.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "score": 1.0, + "content": "Interestingly, at 30-shot, in-context selection with 5 examples outperforms directly adding all 30", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 208, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 506, + 223 + ], + "score": 1.0, + "content": "shots into the prompt. This is showing that in-context selection can help the model utilize a larger", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 220, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 233 + ], + "score": 1.0, + "content": "number noisy video examples. Nevertheless, we still observe that the benefit of adding more shots", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "score": 1.0, + "content": "saturated at around 20 to 30 shots, even if with in-context selection. we view this as a remaining", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 241, + 409, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 409, + 255 + ], + "score": 1.0, + "content": "challenging on how to make language models benefit from longer contexts.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 107, + 275, + 345, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 274, + 346, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 346, + 290 + ], + "score": 1.0, + "content": "5 Conclusions, Limitations and Future Work", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 304, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 105, + 303, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 317 + ], + "score": 1.0, + "content": "This paper proposes VidIL, a few-shot Video-language Learner via Image and Language models.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "It demonstrates the strong ability of large-scale language models on performing video-to-text tasks", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 327, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 505, + 338 + ], + "score": 1.0, + "content": "when frame features are provided as unified text representations using image-language models. We", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 337, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 505, + 349 + ], + "score": 1.0, + "content": "propose a temporal order aware prompt by decomposing videos into a hierarchical structure, which is", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 346, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 506, + 362 + ], + "score": 1.0, + "content": "able to plug in multiple levels of frame features, along with speech transcripts. Without pretraining", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 358, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 372 + ], + "score": 1.0, + "content": "on videos, our model outperforms vision-language models learned from large-scale video datasets", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 369, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 506, + 383 + ], + "score": 1.0, + "content": "on a variety of few-shot tasks, such as domain-specific captioning, question answering, and future", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "event prediction. One limitation of using unified textual representation is that we might lose low-level", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "score": 1.0, + "content": "visual features which can be essential for some specific tasks, such as fine-grained spatial visual", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "question answering. We also observe that current video-language benchmarks rarely require explicit", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 413, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 505, + 426 + ], + "score": 1.0, + "content": "temporal tracking on the frames and visual tokens. Future work will focus on leveraging large-scale", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "score": 1.0, + "content": "language models for learning script knowledge from long videos where temporal dynamics are better", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 435, + 158, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 158, + 448 + ], + "score": 1.0, + "content": "emphasized.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 107, + 468, + 207, + 482 + ], + "lines": [ + { + "bbox": [ + 105, + 467, + 209, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 209, + 484 + ], + "score": 1.0, + "content": "6 Broader Impact", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 497, + 505, + 574 + ], + "lines": [ + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "An open-domain few-shot video-language learner has a wide range of beneficial applications for", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "score": 1.0, + "content": "society, such as automatically detecting violent or mature content in videos and helping people with", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "score": 1.0, + "content": "vision impairment understand videos. However, since the language model is pretrained on massive", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 531, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 542 + ], + "score": 1.0, + "content": "internet-scale text data, there might be unexpected output that can have potential negative impact", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "score": 1.0, + "content": "on the society, such as bias against people of a certain gender, race or sexuality. Future work and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "score": 1.0, + "content": "dedicated collaboration from the community are needed to alleviate the potential negative societal", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 563, + 241, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 241, + 576 + ], + "score": 1.0, + "content": "impact of large language models.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34 + }, + { + "type": "title", + "bbox": [ + 108, + 596, + 207, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 595, + 208, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 208, + 613 + ], + "score": 1.0, + "content": "Acknowledgements", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 625, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "score": 1.0, + "content": "We thank the anonymous reviewers helpful suggestions. This research is based upon work supported", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 636, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 505, + 648 + ], + "score": 1.0, + "content": "in part by U.S. DARPA AIDA Program No. FA8750-18-2-0014 and U.S. DARPA KAIROS Program", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 646, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 506, + 659 + ], + "score": 1.0, + "content": "Nos. FA8750-19-2-1004. The views and conclusions contained herein are those of the authors and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 657, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 506, + 671 + ], + "score": 1.0, + "content": "should not be interpreted as necessarily representing the official policies, either expressed or implied,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "of DARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 680, + 450, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 450, + 693 + ], + "score": 1.0, + "content": "reprints for governmental purposes notwithstanding any copyright annotation therein.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 712, + 363, + 722 + ], + "lines": [ + { + "bbox": [ + 120, + 711, + 365, + 723 + ], + "spans": [ + { + "bbox": [ + 120, + 711, + 365, + 723 + ], + "score": 1.0, + "content": "9we use the frame caption and visual tokens from the middle frame.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 149 + ], + "lines": [], + "index": 3, + "bbox_fs": [ + 105, + 72, + 506, + 151 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 253 + ], + "lines": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "In Table 7, we first show that, with the same context length, namely, 5 in-context examples, in-context", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 165, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 177 + ], + "score": 1.0, + "content": "example selection significantly increases the performance as well as the robustness. At 10-shot, and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 177, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 506, + 190 + ], + "score": 1.0, + "content": "20-shot, directly fitting more shots into the prompt results in better performance. In-context selection", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 188, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 188, + 506, + 199 + ], + "score": 1.0, + "content": "achieves slightly lower performance but with significantly better efficiency due to shorter context.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "score": 1.0, + "content": "Interestingly, at 30-shot, in-context selection with 5 examples outperforms directly adding all 30", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 208, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 506, + 223 + ], + "score": 1.0, + "content": "shots into the prompt. This is showing that in-context selection can help the model utilize a larger", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 220, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 233 + ], + "score": 1.0, + "content": "number noisy video examples. Nevertheless, we still observe that the benefit of adding more shots", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "score": 1.0, + "content": "saturated at around 20 to 30 shots, even if with in-context selection. we view this as a remaining", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 241, + 409, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 409, + 255 + ], + "score": 1.0, + "content": "challenging on how to make language models benefit from longer contexts.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 154, + 506, + 255 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 275, + 345, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 274, + 346, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 346, + 290 + ], + "score": 1.0, + "content": "5 Conclusions, Limitations and Future Work", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 304, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 105, + 303, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 317 + ], + "score": 1.0, + "content": "This paper proposes VidIL, a few-shot Video-language Learner via Image and Language models.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "It demonstrates the strong ability of large-scale language models on performing video-to-text tasks", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 327, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 505, + 338 + ], + "score": 1.0, + "content": "when frame features are provided as unified text representations using image-language models. We", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 337, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 505, + 349 + ], + "score": 1.0, + "content": "propose a temporal order aware prompt by decomposing videos into a hierarchical structure, which is", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 346, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 506, + 362 + ], + "score": 1.0, + "content": "able to plug in multiple levels of frame features, along with speech transcripts. Without pretraining", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 358, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 372 + ], + "score": 1.0, + "content": "on videos, our model outperforms vision-language models learned from large-scale video datasets", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 369, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 506, + 383 + ], + "score": 1.0, + "content": "on a variety of few-shot tasks, such as domain-specific captioning, question answering, and future", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "event prediction. One limitation of using unified textual representation is that we might lose low-level", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "score": 1.0, + "content": "visual features which can be essential for some specific tasks, such as fine-grained spatial visual", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "question answering. We also observe that current video-language benchmarks rarely require explicit", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 413, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 505, + 426 + ], + "score": 1.0, + "content": "temporal tracking on the frames and visual tokens. Future work will focus on leveraging large-scale", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "score": 1.0, + "content": "language models for learning script knowledge from long videos where temporal dynamics are better", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 435, + 158, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 158, + 448 + ], + "score": 1.0, + "content": "emphasized.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 303, + 506, + 448 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 468, + 207, + 482 + ], + "lines": [ + { + "bbox": [ + 105, + 467, + 209, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 209, + 484 + ], + "score": 1.0, + "content": "6 Broader Impact", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 497, + 505, + 574 + ], + "lines": [ + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "An open-domain few-shot video-language learner has a wide range of beneficial applications for", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "score": 1.0, + "content": "society, such as automatically detecting violent or mature content in videos and helping people with", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "score": 1.0, + "content": "vision impairment understand videos. However, since the language model is pretrained on massive", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 531, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 542 + ], + "score": 1.0, + "content": "internet-scale text data, there might be unexpected output that can have potential negative impact", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "score": 1.0, + "content": "on the society, such as bias against people of a certain gender, race or sexuality. Future work and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "score": 1.0, + "content": "dedicated collaboration from the community are needed to alleviate the potential negative societal", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 563, + 241, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 241, + 576 + ], + "score": 1.0, + "content": "impact of large language models.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 497, + 506, + 576 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 596, + 207, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 595, + 208, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 208, + 613 + ], + "score": 1.0, + "content": "Acknowledgements", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 625, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "score": 1.0, + "content": "We thank the anonymous reviewers helpful suggestions. This research is based upon work supported", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 636, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 505, + 648 + ], + "score": 1.0, + "content": "in part by U.S. DARPA AIDA Program No. FA8750-18-2-0014 and U.S. DARPA KAIROS Program", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 646, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 506, + 659 + ], + "score": 1.0, + "content": "Nos. FA8750-19-2-1004. The views and conclusions contained herein are those of the authors and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 657, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 506, + 671 + ], + "score": 1.0, + "content": "should not be interpreted as necessarily representing the official policies, either expressed or implied,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "of DARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 680, + 450, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 450, + 693 + ], + "score": 1.0, + "content": "reprints for governmental purposes notwithstanding any copyright annotation therein.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 624, + 506, + 693 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 52, + 507, + 724 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 85, + 507, + 103 + ], + "spans": [ + { + "bbox": [ + 107, + 85, + 507, + 103 + ], + "score": 1.0, + "content": "[1] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 124, + 97, + 506, + 112 + ], + "spans": [ + { + "bbox": [ + 124, + 97, + 506, + 112 + ], + "score": 1.0, + "content": "Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. Advances in", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 109, + 317, + 121 + ], + "spans": [ + { + "bbox": [ + 126, + 109, + 317, + 121 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, 34, 2021. 3", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 109, + 125, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 109, + 125, + 506, + 139 + ], + "score": 1.0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 125, + 136, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 125, + 136, + 505, + 147 + ], + "score": 1.0, + "content": "Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 146, + 329, + 158 + ], + "spans": [ + { + "bbox": [ + 126, + 146, + 329, + 158 + ], + "score": 1.0, + "content": "learning. ArXiv preprint, abs/2204.14198, 2022. 3, 7, 8", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 164, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 110, + 164, + 506, + 175 + ], + "score": 1.0, + "content": "[3] Jean-Baptiste Alayrac, Adrià Recasens, Rosalia Schneider, Relja Arandjelovic, Jason Ramapuram, Jef-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 125, + 173, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 125, + 173, + 505, + 185 + ], + "score": 1.0, + "content": "frey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-supervised multimodal", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 126, + 183, + 504, + 195 + ], + "spans": [ + { + "bbox": [ + 126, + 183, + 504, + 195 + ], + "score": 1.0, + "content": "versatile networks. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 125, + 193, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 125, + 193, + 506, + 205 + ], + "score": 1.0, + "content": "Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 125, + 203, + 491, + 215 + ], + "spans": [ + { + "bbox": [ + 125, + 203, + 491, + 215 + ], + "score": 1.0, + "content": "Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 3", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 110, + 220, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 110, + 220, + 506, + 232 + ], + "score": 1.0, + "content": "[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 125, + 230, + 504, + 242 + ], + "spans": [ + { + "bbox": [ + 125, + 230, + 504, + 242 + ], + "score": 1.0, + "content": "and Devi Parikh. VQA: visual question answering. In 2015 IEEE International Conference on Computer", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 125, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 125, + 240, + 506, + 254 + ], + "score": 1.0, + "content": "Vision, ICCV 2015, Santiago, Chile, December 7-13, 2015, pages 2425–2433. IEEE Computer Society,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 126, + 250, + 167, + 262 + ], + "spans": [ + { + "bbox": [ + 126, + 250, + 167, + 262 + ], + "score": 1.0, + "content": "2015. 7, 8", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 110, + 268, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 110, + 268, + 506, + 279 + ], + "score": 1.0, + "content": "[5] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 126, + 278, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 126, + 278, + 505, + 290 + ], + "score": 1.0, + "content": "correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 125, + 286, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 125, + 286, + 506, + 300 + ], + "score": 1.0, + "content": "Evaluation Measures for Machine Translation and/or Summarization, pages 65–72, Ann Arbor, Michigan,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 124, + 295, + 317, + 310 + ], + "spans": [ + { + "bbox": [ + 124, + 295, + 317, + 310 + ], + "score": 1.0, + "content": "2005. Association for Computational Linguistics. 6", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 110, + 315, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 110, + 315, + 506, + 326 + ], + "score": 1.0, + "content": "[6] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 125, + 325, + 506, + 337 + ], + "spans": [ + { + "bbox": [ + 125, + 325, + 506, + 337 + ], + "score": 1.0, + "content": "Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 125, + 335, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 125, + 335, + 506, + 347 + ], + "score": 1.0, + "content": "Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 126, + 344, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 126, + 344, + 506, + 356 + ], + "score": 1.0, + "content": "Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 124, + 353, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 124, + 353, + 506, + 368 + ], + "score": 1.0, + "content": "Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 125, + 364, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 125, + 364, + 505, + 376 + ], + "score": 1.0, + "content": "models are few-shot learners. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 125, + 375, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 125, + 375, + 506, + 386 + ], + "score": 1.0, + "content": "Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 125, + 385, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 125, + 385, + 506, + 397 + ], + "score": 1.0, + "content": "Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 125, + 393, + 176, + 407 + ], + "spans": [ + { + "bbox": [ + 125, + 393, + 176, + 407 + ], + "score": 1.0, + "content": "2020. 1, 3, 5", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 110, + 411, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 110, + 411, + 506, + 423 + ], + "score": 1.0, + "content": "[7] David Chen and William Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 125, + 422, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 125, + 422, + 506, + 434 + ], + "score": 1.0, + "content": "the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 124, + 430, + 463, + 444 + ], + "spans": [ + { + "bbox": [ + 124, + 430, + 463, + 444 + ], + "score": 1.0, + "content": "pages 190–200, Portland, Oregon, USA, 2011. Association for Computational Linguistics. 6", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 109, + 447, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 109, + 447, + 506, + 461 + ], + "score": 1.0, + "content": "[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 124, + 457, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 124, + 457, + 506, + 472 + ], + "score": 1.0, + "content": "Liu. Uniter: Universal image-text representation learning. In European conference on computer vision,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 124, + 468, + 252, + 481 + ], + "spans": [ + { + "bbox": [ + 124, + 468, + 252, + 481 + ], + "score": 1.0, + "content": "pages 104–120. Springer, 2020. 2", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 109, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 109, + 484, + 506, + 498 + ], + "score": 1.0, + "content": "[9] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 125, + 496, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 125, + 496, + 505, + 508 + ], + "score": 1.0, + "content": "In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 124, + 504, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 124, + 504, + 506, + 519 + ], + "score": 1.0, + "content": "Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 124, + 515, + 293, + 529 + ], + "spans": [ + { + "bbox": [ + 124, + 515, + 293, + 529 + ], + "score": 1.0, + "content": "Research, pages 1931–1942. PMLR, 2021. 3", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 532, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 104, + 532, + 506, + 545 + ], + "score": 1.0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. Imagenet: A large-scale hierarchical", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 124, + 542, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 124, + 542, + 506, + 556 + ], + "score": 1.0, + "content": "image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 124, + 552, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 124, + 552, + 505, + 566 + ], + "score": 1.0, + "content": "(CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248–255. IEEE Computer Society, 2009. 4", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "score": 1.0, + "content": "[11] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 125, + 580, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 125, + 580, + 295, + 592 + ], + "score": 1.0, + "content": "clip. ArXiv preprint, abs/2106.11097, 2021. 3", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 597, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 506, + 609 + ], + "score": 1.0, + "content": "[12] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 124, + 606, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 124, + 606, + 506, + 620 + ], + "score": 1.0, + "content": "recognition. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 125, + 617, + 398, + 629 + ], + "spans": [ + { + "bbox": [ + 125, + 617, + 398, + 629 + ], + "score": 1.0, + "content": "(South), October 27 - November 2, 2019, pages 6201–6210. IEEE, 2019. 3", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 634, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 506, + 646 + ], + "score": 1.0, + "content": "[13] Tianyu Gao, Adam Fisch, and Danqi Chen. Making pre-trained language models better few-shot learners.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 125, + 645, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 125, + 645, + 506, + 656 + ], + "score": 1.0, + "content": "In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 124, + 652, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 124, + 652, + 506, + 668 + ], + "score": 1.0, + "content": "11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 124, + 662, + 390, + 677 + ], + "spans": [ + { + "bbox": [ + 124, + 662, + 390, + 677 + ], + "score": 1.0, + "content": "3816–3830, Online, 2021. Association for Computational Linguistics. 2", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 681, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 506, + 693 + ], + "score": 1.0, + "content": "[14] Matt Gardner, Joel Grus, Mark Neumann, Oyvind Tafjord, Pradeep Dasigi, Nelson F. Liu, Matthew Peters,", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 124, + 691, + 507, + 705 + ], + "spans": [ + { + "bbox": [ + 124, + 691, + 507, + 705 + ], + "score": 1.0, + "content": "Michael Schmitz, and Luke Zettlemoyer. AllenNLP: A deep semantic natural language processing platform.", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 124, + 700, + 507, + 714 + ], + "spans": [ + { + "bbox": [ + 124, + 700, + 507, + 714 + ], + "score": 1.0, + "content": "In Proceedings of Workshop for NLP Open Source Software (NLP-OSS), pages 1–6, Melbourne, Australia,", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 124, + 709, + 316, + 724 + ], + "spans": [ + { + "bbox": [ + 124, + 709, + 316, + 724 + ], + "score": 1.0, + "content": "2018. Association for Computational Linguistics. 2", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 27 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 310, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 52, + 507, + 724 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 107, + 85, + 507, + 103 + ], + "spans": [ + { + "bbox": [ + 107, + 85, + 507, + 103 + ], + "score": 1.0, + "content": "[1] Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong.", + "type": "text" + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 97, + 506, + 112 + ], + "spans": [ + { + "bbox": [ + 124, + 97, + 506, + 112 + ], + "score": 1.0, + "content": "Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. Advances in", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 109, + 317, + 121 + ], + "spans": [ + { + "bbox": [ + 126, + 109, + 317, + 121 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, 34, 2021. 3", + "type": "text" + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 125, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 109, + 125, + 506, + 139 + ], + "score": 1.0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc,", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 136, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 125, + 136, + 505, + 147 + ], + "score": 1.0, + "content": "Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 146, + 329, + 158 + ], + "spans": [ + { + "bbox": [ + 126, + 146, + 329, + 158 + ], + "score": 1.0, + "content": "learning. ArXiv preprint, abs/2204.14198, 2022. 3, 7, 8", + "type": "text" + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 164, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 110, + 164, + 506, + 175 + ], + "score": 1.0, + "content": "[3] Jean-Baptiste Alayrac, Adrià Recasens, Rosalia Schneider, Relja Arandjelovic, Jason Ramapuram, Jef-", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 173, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 125, + 173, + 505, + 185 + ], + "score": 1.0, + "content": "frey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-supervised multimodal", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 126, + 183, + 504, + 195 + ], + "spans": [ + { + "bbox": [ + 126, + 183, + 504, + 195 + ], + "score": 1.0, + "content": "versatile networks. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 125, + 193, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 125, + 193, + 506, + 205 + ], + "score": 1.0, + "content": "Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 125, + 203, + 491, + 215 + ], + "spans": [ + { + "bbox": [ + 125, + 203, + 491, + 215 + ], + "score": 1.0, + "content": "Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 3", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 110, + 220, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 110, + 220, + 506, + 232 + ], + "score": 1.0, + "content": "[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick,", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 230, + 504, + 242 + ], + "spans": [ + { + "bbox": [ + 125, + 230, + 504, + 242 + ], + "score": 1.0, + "content": "and Devi Parikh. VQA: visual question answering. In 2015 IEEE International Conference on Computer", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 125, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 125, + 240, + 506, + 254 + ], + "score": 1.0, + "content": "Vision, ICCV 2015, Santiago, Chile, December 7-13, 2015, pages 2425–2433. IEEE Computer Society,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 126, + 250, + 167, + 262 + ], + "spans": [ + { + "bbox": [ + 126, + 250, + 167, + 262 + ], + "score": 1.0, + "content": "2015. 7, 8", + "type": "text" + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 268, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 110, + 268, + 506, + 279 + ], + "score": 1.0, + "content": "[5] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 278, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 126, + 278, + 505, + 290 + ], + "score": 1.0, + "content": "correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 125, + 286, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 125, + 286, + 506, + 300 + ], + "score": 1.0, + "content": "Evaluation Measures for Machine Translation and/or Summarization, pages 65–72, Ann Arbor, Michigan,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 124, + 295, + 317, + 310 + ], + "spans": [ + { + "bbox": [ + 124, + 295, + 317, + 310 + ], + "score": 1.0, + "content": "2005. Association for Computational Linguistics. 6", + "type": "text" + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 315, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 110, + 315, + 506, + 326 + ], + "score": 1.0, + "content": "[6] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 325, + 506, + 337 + ], + "spans": [ + { + "bbox": [ + 125, + 325, + 506, + 337 + ], + "score": 1.0, + "content": "Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 125, + 335, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 125, + 335, + 506, + 347 + ], + "score": 1.0, + "content": "Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 126, + 344, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 126, + 344, + 506, + 356 + ], + "score": 1.0, + "content": "Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 124, + 353, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 124, + 353, + 506, + 368 + ], + "score": 1.0, + "content": "Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 125, + 364, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 125, + 364, + 505, + 376 + ], + "score": 1.0, + "content": "models are few-shot learners. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 125, + 375, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 125, + 375, + 506, + 386 + ], + "score": 1.0, + "content": "Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 125, + 385, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 125, + 385, + 506, + 397 + ], + "score": 1.0, + "content": "Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 125, + 393, + 176, + 407 + ], + "spans": [ + { + "bbox": [ + 125, + 393, + 176, + 407 + ], + "score": 1.0, + "content": "2020. 1, 3, 5", + "type": "text" + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 411, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 110, + 411, + 506, + 423 + ], + "score": 1.0, + "content": "[7] David Chen and William Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 422, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 125, + 422, + 506, + 434 + ], + "score": 1.0, + "content": "the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 124, + 430, + 463, + 444 + ], + "spans": [ + { + "bbox": [ + 124, + 430, + 463, + 444 + ], + "score": 1.0, + "content": "pages 190–200, Portland, Oregon, USA, 2011. Association for Computational Linguistics. 6", + "type": "text" + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 447, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 109, + 447, + 506, + 461 + ], + "score": 1.0, + "content": "[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 457, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 124, + 457, + 506, + 472 + ], + "score": 1.0, + "content": "Liu. Uniter: Universal image-text representation learning. In European conference on computer vision,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 124, + 468, + 252, + 481 + ], + "spans": [ + { + "bbox": [ + 124, + 468, + 252, + 481 + ], + "score": 1.0, + "content": "pages 104–120. Springer, 2020. 2", + "type": "text" + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 109, + 484, + 506, + 498 + ], + "score": 1.0, + "content": "[9] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation.", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 496, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 125, + 496, + 505, + 508 + ], + "score": 1.0, + "content": "In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 124, + 504, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 124, + 504, + 506, + 519 + ], + "score": 1.0, + "content": "Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 124, + 515, + 293, + 529 + ], + "spans": [ + { + "bbox": [ + 124, + 515, + 293, + 529 + ], + "score": 1.0, + "content": "Research, pages 1931–1942. PMLR, 2021. 3", + "type": "text" + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 532, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 104, + 532, + 506, + 545 + ], + "score": 1.0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. Imagenet: A large-scale hierarchical", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 542, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 124, + 542, + 506, + 556 + ], + "score": 1.0, + "content": "image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 124, + 552, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 124, + 552, + 505, + 566 + ], + "score": 1.0, + "content": "(CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248–255. IEEE Computer Society, 2009. 4", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "score": 1.0, + "content": "[11] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image", + "type": "text" + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 580, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 125, + 580, + 295, + 592 + ], + "score": 1.0, + "content": "clip. ArXiv preprint, abs/2106.11097, 2021. 3", + "type": "text" + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 597, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 506, + 609 + ], + "score": 1.0, + "content": "[12] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 606, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 124, + 606, + 506, + 620 + ], + "score": 1.0, + "content": "recognition. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 125, + 617, + 398, + 629 + ], + "spans": [ + { + "bbox": [ + 125, + 617, + 398, + 629 + ], + "score": 1.0, + "content": "(South), October 27 - November 2, 2019, pages 6201–6210. IEEE, 2019. 3", + "type": "text" + } + ], + "index": 46, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 634, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 506, + 646 + ], + "score": 1.0, + "content": "[13] Tianyu Gao, Adam Fisch, and Danqi Chen. Making pre-trained language models better few-shot learners.", + "type": "text" + } + ], + "index": 47, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 645, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 125, + 645, + 506, + 656 + ], + "score": 1.0, + "content": "In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 124, + 652, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 124, + 652, + 506, + 668 + ], + "score": 1.0, + "content": "11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 124, + 662, + 390, + 677 + ], + "spans": [ + { + "bbox": [ + 124, + 662, + 390, + 677 + ], + "score": 1.0, + "content": "3816–3830, Online, 2021. Association for Computational Linguistics. 2", + "type": "text" + } + ], + "index": 50, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 681, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 506, + 693 + ], + "score": 1.0, + "content": "[14] Matt Gardner, Joel Grus, Mark Neumann, Oyvind Tafjord, Pradeep Dasigi, Nelson F. Liu, Matthew Peters,", + "type": "text" + } + ], + "index": 51, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 691, + 507, + 705 + ], + "spans": [ + { + "bbox": [ + 124, + 691, + 507, + 705 + ], + "score": 1.0, + "content": "Michael Schmitz, and Luke Zettlemoyer. AllenNLP: A deep semantic natural language processing platform.", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 124, + 700, + 507, + 714 + ], + "spans": [ + { + "bbox": [ + 124, + 700, + 507, + 714 + ], + "score": 1.0, + "content": "In Proceedings of Workshop for NLP Open Source Software (NLP-OSS), pages 1–6, Melbourne, Australia,", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 124, + 709, + 316, + 724 + ], + "spans": [ + { + "bbox": [ + 124, + 709, + 316, + 724 + ], + "score": 1.0, + "content": "2018. Association for Computational Linguistics. 2", + "type": "text" + } + ], + "index": 54, + "is_list_end_line": true + } + ], + "index": 27, + "bbox_fs": [ + 104, + 70, + 507, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 73, + 506, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 73, + 506, + 84 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 506, + 84 + ], + "score": 1.0, + "content": "[15] Ting-Hao Kenneth Huang, Francis Ferraro, Nasrin Mostafazadeh, Ishan Misra, Aishwarya Agrawal, Jacob", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 83, + 506, + 94 + ], + "spans": [ + { + "bbox": [ + 125, + 83, + 506, + 94 + ], + "score": 1.0, + "content": "Devlin, Ross Girshick, Xiaodong He, Pushmeet Kohli, Dhruv Batra, C. Lawrence Zitnick, Devi Parikh,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 125, + 93, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 125, + 93, + 506, + 105 + ], + "score": 1.0, + "content": "Lucy Vanderwende, Michel Galley, and Margaret Mitchell. Visual storytelling. In Proceedings of the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 125, + 102, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 125, + 102, + 505, + 115 + ], + "score": 1.0, + "content": "2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 114, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 125, + 114, + 505, + 124 + ], + "score": 1.0, + "content": "Language Technologies, pages 1233–1239, San Diego, California, 2016. Association for Computational", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 124, + 123, + 180, + 135 + ], + "spans": [ + { + "bbox": [ + 124, + 123, + 180, + 135 + ], + "score": 1.0, + "content": "Linguistics. 2", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 105, + 141, + 502, + 172 + ], + "lines": [ + { + "bbox": [ + 105, + 140, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 505, + 153 + ], + "score": 1.0, + "content": "[16] Zhicheng Huang, Zhaoyang Zeng, Yupan Huang, Bei Liu, Dongmei Fu, and Jianlong Fu. Seeing out of the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 125, + 150, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 125, + 150, + 505, + 163 + ], + "score": 1.0, + "content": "box: End-to-end pre-training for vision-language representation learning. In Proceedings of the IEEE/CVF", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 125, + 160, + 444, + 173 + ], + "spans": [ + { + "bbox": [ + 125, + 160, + 444, + 173 + ], + "score": 1.0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 12976–12985, 2021. 2", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 178, + 505, + 229 + ], + "lines": [ + { + "bbox": [ + 105, + 177, + 506, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 506, + 191 + ], + "score": 1.0, + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 125, + 188, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 125, + 188, + 505, + 200 + ], + "score": 1.0, + "content": "Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 126, + 199, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 126, + 199, + 506, + 211 + ], + "score": 1.0, + "content": "supervision. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 126, + 208, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 126, + 208, + 505, + 220 + ], + "score": 1.0, + "content": "on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 126, + 218, + 326, + 230 + ], + "spans": [ + { + "bbox": [ + 126, + 218, + 326, + 230 + ], + "score": 1.0, + "content": "Learning Research, pages 4904–4916. PMLR, 2021. 2", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 236, + 505, + 277 + ], + "lines": [ + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "score": 1.0, + "content": "[18] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 126, + 247, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 126, + 247, + 505, + 258 + ], + "score": 1.0, + "content": "or region supervision. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 125, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 125, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 125, + 266, + 370, + 277 + ], + "spans": [ + { + "bbox": [ + 125, + 266, + 370, + 277 + ], + "score": 1.0, + "content": "of Machine Learning Research, pages 5583–5594. PMLR, 2021. 2", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 284, + 504, + 324 + ], + "lines": [ + { + "bbox": [ + 106, + 284, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 506, + 296 + ], + "score": 1.0, + "content": "[19] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 125, + 293, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 125, + 293, + 505, + 306 + ], + "score": 1.0, + "content": "Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 126, + 304, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 126, + 304, + 506, + 316 + ], + "score": 1.0, + "content": "using crowdsourced dense image annotations. International journal of computer vision, 123(1):32–73,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 126, + 314, + 158, + 325 + ], + "spans": [ + { + "bbox": [ + 126, + 314, + 158, + 325 + ], + "score": 1.0, + "content": "2017. 4", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 105, + 331, + 504, + 363 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 344 + ], + "score": 1.0, + "content": "[20] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Ka-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 126, + 342, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 126, + 342, + 506, + 353 + ], + "score": 1.0, + "content": "mali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 125, + 352, + 333, + 363 + ], + "spans": [ + { + "bbox": [ + 125, + 352, + 333, + 363 + ], + "score": 1.0, + "content": "Journal of Computer Vision, 128(7):1956–1981, 2020. 4", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 105, + 369, + 504, + 390 + ], + "lines": [ + { + "bbox": [ + 106, + 369, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 505, + 380 + ], + "score": 1.0, + "content": "[21] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L. Berg, Mohit Bansal, and Jingjing Liu. Less is more:", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 125, + 379, + 427, + 391 + ], + "spans": [ + { + "bbox": [ + 125, + 379, + 427, + 391 + ], + "score": 1.0, + "content": "Clipbert for video-and-language learningvia sparse sampling. In CVPR, 2021. 1, 3", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 504, + 428 + ], + "lines": [ + { + "bbox": [ + 106, + 397, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 506, + 410 + ], + "score": 1.0, + "content": "[22] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara Berg. TVQA: Localized, compositional video question", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 125, + 407, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 125, + 407, + 506, + 420 + ], + "score": 1.0, + "content": "answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 125, + 417, + 453, + 429 + ], + "spans": [ + { + "bbox": [ + 125, + 417, + 453, + 429 + ], + "score": 1.0, + "content": "pages 1369–1379, Brussels, Belgium, 2018. Association for Computational Linguistics. 3", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 435, + 505, + 475 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 507, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 507, + 448 + ], + "score": 1.0, + "content": "[23] Jie Lei, Licheng Yu, Tamara Berg, and Mohit Bansal. What is more likely to happen next? video-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 126, + 446, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 126, + 446, + 505, + 456 + ], + "score": 1.0, + "content": "and-language future event prediction. In Proceedings of the 2020 Conference on Empirical Methods in", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 126, + 455, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 126, + 455, + 505, + 466 + ], + "score": 1.0, + "content": "Natural Language Processing (EMNLP), pages 8769–8784, Online, 2020. Association for Computational", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 126, + 465, + 198, + 477 + ], + "spans": [ + { + "bbox": [ + 126, + 465, + 198, + 477 + ], + "score": 1.0, + "content": "Linguistics. 1, 6, 8", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 106, + 482, + 505, + 533 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 507, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 507, + 495 + ], + "score": 1.0, + "content": "[24] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 126, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 126, + 492, + 505, + 505 + ], + "score": 1.0, + "content": "Veselin Stoyanov, and Luke Zettlemoyer. BART: Denoising sequence-to-sequence pre-training for natural", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 124, + 502, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 124, + 502, + 506, + 515 + ], + "score": 1.0, + "content": "language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 124, + 512, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 124, + 512, + 506, + 525 + ], + "score": 1.0, + "content": "Association for Computational Linguistics, pages 7871–7880, Online, 2020. Association for Computational", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 124, + 522, + 179, + 535 + ], + "spans": [ + { + "bbox": [ + 124, + 522, + 179, + 535 + ], + "score": 1.0, + "content": "Linguistics. 1", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 105, + 540, + 504, + 561 + ], + "lines": [ + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "score": 1.0, + "content": "[25] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven C.H. Hoi. Align and prompt:", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 126, + 550, + 412, + 562 + ], + "spans": [ + { + "bbox": [ + 126, + 550, + 412, + 562 + ], + "score": 1.0, + "content": "Video-and-language pre-training with entity prompts. In arxiv, 2021. 1, 3, 8, 9", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 106, + 568, + 504, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 581 + ], + "score": 1.0, + "content": "[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 125, + 578, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 125, + 578, + 506, + 590 + ], + "score": 1.0, + "content": "for unified vision-language understanding and generation. ArXiv preprint, abs/2201.12086, 2022. 2, 3, 6,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 124, + 587, + 144, + 600 + ], + "spans": [ + { + "bbox": [ + 124, + 587, + 144, + 600 + ], + "score": 1.0, + "content": "7, 8", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 106, + 605, + 504, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 619 + ], + "score": 1.0, + "content": "[27] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 124, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 124, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 126, + 626, + 327, + 638 + ], + "spans": [ + { + "bbox": [ + 126, + 626, + 327, + 638 + ], + "score": 1.0, + "content": "in Neural Information Processing Systems, 34, 2021. 2", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 106, + 643, + 504, + 684 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "score": 1.0, + "content": "[28] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. HERO: Hierarchical encoder", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 125, + 652, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 125, + 652, + 506, + 666 + ], + "score": 1.0, + "content": "for Video+Language omni-representation pre-training. In Proceedings of the 2020 Conference on Empirical", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 125, + 662, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 125, + 662, + 506, + 676 + ], + "score": 1.0, + "content": "Methods in Natural Language Processing (EMNLP), pages 2046–2065, Online, 2020. Association for", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 126, + 673, + 244, + 685 + ], + "spans": [ + { + "bbox": [ + 126, + 673, + 244, + 685 + ], + "score": 1.0, + "content": "Computational Linguistics. 1, 3", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48.5 + }, + { + "type": "text", + "bbox": [ + 107, + 691, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 690, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 505, + 704 + ], + "score": 1.0, + "content": "[29] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 125, + 701, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 125, + 701, + 505, + 713 + ], + "score": 1.0, + "content": "Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 126, + 711, + 408, + 724 + ], + "spans": [ + { + "bbox": [ + 126, + 711, + 408, + 724 + ], + "score": 1.0, + "content": "European Conference on Computer Vision, pages 121–137. Springer, 2020. 2", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 52 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 73, + 506, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 73, + 506, + 84 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 506, + 84 + ], + "score": 1.0, + "content": "[15] Ting-Hao Kenneth Huang, Francis Ferraro, Nasrin Mostafazadeh, Ishan Misra, Aishwarya Agrawal, Jacob", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 83, + 506, + 94 + ], + "spans": [ + { + "bbox": [ + 125, + 83, + 506, + 94 + ], + "score": 1.0, + "content": "Devlin, Ross Girshick, Xiaodong He, Pushmeet Kohli, Dhruv Batra, C. Lawrence Zitnick, Devi Parikh,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 125, + 93, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 125, + 93, + 506, + 105 + ], + "score": 1.0, + "content": "Lucy Vanderwende, Michel Galley, and Margaret Mitchell. Visual storytelling. In Proceedings of the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 125, + 102, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 125, + 102, + 505, + 115 + ], + "score": 1.0, + "content": "2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 114, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 125, + 114, + 505, + 124 + ], + "score": 1.0, + "content": "Language Technologies, pages 1233–1239, San Diego, California, 2016. Association for Computational", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 124, + 123, + 180, + 135 + ], + "spans": [ + { + "bbox": [ + 124, + 123, + 180, + 135 + ], + "score": 1.0, + "content": "Linguistics. 2", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 73, + 506, + 135 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 141, + 502, + 172 + ], + "lines": [ + { + "bbox": [ + 105, + 140, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 505, + 153 + ], + "score": 1.0, + "content": "[16] Zhicheng Huang, Zhaoyang Zeng, Yupan Huang, Bei Liu, Dongmei Fu, and Jianlong Fu. Seeing out of the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 125, + 150, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 125, + 150, + 505, + 163 + ], + "score": 1.0, + "content": "box: End-to-end pre-training for vision-language representation learning. In Proceedings of the IEEE/CVF", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 125, + 160, + 444, + 173 + ], + "spans": [ + { + "bbox": [ + 125, + 160, + 444, + 173 + ], + "score": 1.0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 12976–12985, 2021. 2", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 140, + 505, + 173 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 178, + 505, + 229 + ], + "lines": [ + { + "bbox": [ + 105, + 177, + 506, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 506, + 191 + ], + "score": 1.0, + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 125, + 188, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 125, + 188, + 505, + 200 + ], + "score": 1.0, + "content": "Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 126, + 199, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 126, + 199, + 506, + 211 + ], + "score": 1.0, + "content": "supervision. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 126, + 208, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 126, + 208, + 505, + 220 + ], + "score": 1.0, + "content": "on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 126, + 218, + 326, + 230 + ], + "spans": [ + { + "bbox": [ + 126, + 218, + 326, + 230 + ], + "score": 1.0, + "content": "Learning Research, pages 4904–4916. PMLR, 2021. 2", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 177, + 506, + 230 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 236, + 505, + 277 + ], + "lines": [ + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "score": 1.0, + "content": "[18] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 126, + 247, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 126, + 247, + 505, + 258 + ], + "score": 1.0, + "content": "or region supervision. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 125, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 125, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 125, + 266, + 370, + 277 + ], + "spans": [ + { + "bbox": [ + 125, + 266, + 370, + 277 + ], + "score": 1.0, + "content": "of Machine Learning Research, pages 5583–5594. PMLR, 2021. 2", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5, + "bbox_fs": [ + 106, + 236, + 506, + 277 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 284, + 504, + 324 + ], + "lines": [ + { + "bbox": [ + 106, + 284, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 506, + 296 + ], + "score": 1.0, + "content": "[19] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 125, + 293, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 125, + 293, + 505, + 306 + ], + "score": 1.0, + "content": "Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 126, + 304, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 126, + 304, + 506, + 316 + ], + "score": 1.0, + "content": "using crowdsourced dense image annotations. International journal of computer vision, 123(1):32–73,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 126, + 314, + 158, + 325 + ], + "spans": [ + { + "bbox": [ + 126, + 314, + 158, + 325 + ], + "score": 1.0, + "content": "2017. 4", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5, + "bbox_fs": [ + 106, + 284, + 506, + 325 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 331, + 504, + 363 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 344 + ], + "score": 1.0, + "content": "[20] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Ka-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 126, + 342, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 126, + 342, + 506, + 353 + ], + "score": 1.0, + "content": "mali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 125, + 352, + 333, + 363 + ], + "spans": [ + { + "bbox": [ + 125, + 352, + 333, + 363 + ], + "score": 1.0, + "content": "Journal of Computer Vision, 128(7):1956–1981, 2020. 4", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 331, + 506, + 363 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 369, + 504, + 390 + ], + "lines": [ + { + "bbox": [ + 106, + 369, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 505, + 380 + ], + "score": 1.0, + "content": "[21] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L. Berg, Mohit Bansal, and Jingjing Liu. Less is more:", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 125, + 379, + 427, + 391 + ], + "spans": [ + { + "bbox": [ + 125, + 379, + 427, + 391 + ], + "score": 1.0, + "content": "Clipbert for video-and-language learningvia sparse sampling. In CVPR, 2021. 1, 3", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 106, + 369, + 505, + 391 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 504, + 428 + ], + "lines": [ + { + "bbox": [ + 106, + 397, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 506, + 410 + ], + "score": 1.0, + "content": "[22] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara Berg. TVQA: Localized, compositional video question", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 125, + 407, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 125, + 407, + 506, + 420 + ], + "score": 1.0, + "content": "answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 125, + 417, + 453, + 429 + ], + "spans": [ + { + "bbox": [ + 125, + 417, + 453, + 429 + ], + "score": 1.0, + "content": "pages 1369–1379, Brussels, Belgium, 2018. Association for Computational Linguistics. 3", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 106, + 397, + 506, + 429 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 435, + 505, + 475 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 507, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 507, + 448 + ], + "score": 1.0, + "content": "[23] Jie Lei, Licheng Yu, Tamara Berg, and Mohit Bansal. What is more likely to happen next? video-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 126, + 446, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 126, + 446, + 505, + 456 + ], + "score": 1.0, + "content": "and-language future event prediction. In Proceedings of the 2020 Conference on Empirical Methods in", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 126, + 455, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 126, + 455, + 505, + 466 + ], + "score": 1.0, + "content": "Natural Language Processing (EMNLP), pages 8769–8784, Online, 2020. Association for Computational", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 126, + 465, + 198, + 477 + ], + "spans": [ + { + "bbox": [ + 126, + 465, + 198, + 477 + ], + "score": 1.0, + "content": "Linguistics. 1, 6, 8", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 434, + 507, + 477 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 482, + 505, + 533 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 507, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 507, + 495 + ], + "score": 1.0, + "content": "[24] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 126, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 126, + 492, + 505, + 505 + ], + "score": 1.0, + "content": "Veselin Stoyanov, and Luke Zettlemoyer. BART: Denoising sequence-to-sequence pre-training for natural", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 124, + 502, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 124, + 502, + 506, + 515 + ], + "score": 1.0, + "content": "language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 124, + 512, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 124, + 512, + 506, + 525 + ], + "score": 1.0, + "content": "Association for Computational Linguistics, pages 7871–7880, Online, 2020. Association for Computational", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 124, + 522, + 179, + 535 + ], + "spans": [ + { + "bbox": [ + 124, + 522, + 179, + 535 + ], + "score": 1.0, + "content": "Linguistics. 1", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 482, + 507, + 535 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 540, + 504, + 561 + ], + "lines": [ + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "score": 1.0, + "content": "[25] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven C.H. Hoi. Align and prompt:", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 126, + 550, + 412, + 562 + ], + "spans": [ + { + "bbox": [ + 126, + 550, + 412, + 562 + ], + "score": 1.0, + "content": "Video-and-language pre-training with entity prompts. In arxiv, 2021. 1, 3, 8, 9", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 539, + 506, + 562 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 568, + 504, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 581 + ], + "score": 1.0, + "content": "[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 125, + 578, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 125, + 578, + 506, + 590 + ], + "score": 1.0, + "content": "for unified vision-language understanding and generation. ArXiv preprint, abs/2201.12086, 2022. 2, 3, 6,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 124, + 587, + 144, + 600 + ], + "spans": [ + { + "bbox": [ + 124, + 587, + 144, + 600 + ], + "score": 1.0, + "content": "7, 8", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 567, + 506, + 600 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 605, + 504, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 619 + ], + "score": 1.0, + "content": "[27] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 124, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 124, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 126, + 626, + 327, + 638 + ], + "spans": [ + { + "bbox": [ + 126, + 626, + 327, + 638 + ], + "score": 1.0, + "content": "in Neural Information Processing Systems, 34, 2021. 2", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 604, + 506, + 638 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 643, + 504, + 684 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "score": 1.0, + "content": "[28] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. HERO: Hierarchical encoder", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 125, + 652, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 125, + 652, + 506, + 666 + ], + "score": 1.0, + "content": "for Video+Language omni-representation pre-training. In Proceedings of the 2020 Conference on Empirical", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 125, + 662, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 125, + 662, + 506, + 676 + ], + "score": 1.0, + "content": "Methods in Natural Language Processing (EMNLP), pages 2046–2065, Online, 2020. Association for", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 126, + 673, + 244, + 685 + ], + "spans": [ + { + "bbox": [ + 126, + 673, + 244, + 685 + ], + "score": 1.0, + "content": "Computational Linguistics. 1, 3", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48.5, + "bbox_fs": [ + 106, + 644, + 506, + 685 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 691, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 690, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 505, + 704 + ], + "score": 1.0, + "content": "[29] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 125, + 701, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 125, + 701, + 505, + 713 + ], + "score": 1.0, + "content": "Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 126, + 711, + 408, + 724 + ], + "spans": [ + { + "bbox": [ + 126, + 711, + 408, + 724 + ], + "score": 1.0, + "content": "European Conference on Computer Vision, pages 121–137. Springer, 2020. 2", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 52, + "bbox_fs": [ + 105, + 690, + 505, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 73, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "score": 1.0, + "content": "[30] Chin-Yew Lin and Franz Josef Och. Automatic evaluation of machine translation quality using longest", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 125, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "common subsequence and skip-bigram statistics. In Proceedings of the 42nd Annual Meeting of the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 125, + 93, + 475, + 105 + ], + "spans": [ + { + "bbox": [ + 125, + 93, + 475, + 105 + ], + "score": 1.0, + "content": "Association for Computational Linguistics (ACL-04), pages 605–612, Barcelona, Spain, 2004. 6", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 503, + 143 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 124 + ], + "score": 1.0, + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 121, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 125, + 121, + 505, + 134 + ], + "score": 1.0, + "content": "and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 132, + 313, + 143 + ], + "spans": [ + { + "bbox": [ + 126, + 132, + 313, + 143 + ], + "score": 1.0, + "content": "computer vision, pages 740–755. Springer, 2014. 6", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 150, + 504, + 181 + ], + "lines": [ + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "score": 1.0, + "content": "[32] Xudong Lin, Gedas Bertasius, Jue Wang, Shih-Fu Chang, Devi Parikh, and Lorenzo Torresani. Vx2text:", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 161, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 126, + 161, + 505, + 171 + ], + "score": 1.0, + "content": "End-to-end learning of video-based text generation from multimodal inputs. In 2021 IEEE/CVF Conference", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 125, + 170, + 454, + 182 + ], + "spans": [ + { + "bbox": [ + 125, + 170, + 454, + 182 + ], + "score": 1.0, + "content": "on Computer Vision and Pattern Recognition (CVPR), pages 7001–7011. IEEE, 2021. 3, 4", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 105, + 189, + 505, + 219 + ], + "lines": [ + { + "bbox": [ + 105, + 188, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 506, + 200 + ], + "score": 1.0, + "content": "[33] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 124, + 199, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 124, + 199, + 506, + 210 + ], + "score": 1.0, + "content": "Learning to recognize procedural activities with distant supervision. In Proceedings of the IEEE/CVF", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 126, + 209, + 443, + 220 + ], + "spans": [ + { + "bbox": [ + 126, + 209, + 443, + 220 + ], + "score": 1.0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 13853–13863, 2022. 3", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 227, + 505, + 257 + ], + "lines": [ + { + "bbox": [ + 105, + 226, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 506, + 240 + ], + "score": 1.0, + "content": "[34] Xudong Lin, Simran Tiwari, Shiyuan Huang, Manling Li, Mike Zheng Shou, Heng Ji, and Shih-Fu Chang.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 125, + 237, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 125, + 237, + 505, + 248 + ], + "score": 1.0, + "content": "Towards fast adaptation of pretrained contrastive models for multi-channel video-language retrieval. arXiv", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 125, + 247, + 258, + 258 + ], + "spans": [ + { + "bbox": [ + 125, + 247, + 258, + 258 + ], + "score": 1.0, + "content": "preprint arXiv:2206.02082, 2022. 1", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 265, + 505, + 316 + ], + "lines": [ + { + "bbox": [ + 106, + 266, + 504, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 504, + 277 + ], + "score": 1.0, + "content": "[35] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 125, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 125, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "representations for vision-and-language tasks. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 125, + 285, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 125, + 285, + 505, + 297 + ], + "score": 1.0, + "content": "Florence d’Alché-Buc, Emily B. Fox, and Roman Garnett, editors, Advances in Neural Information", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 126, + 296, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 126, + 296, + 505, + 307 + ], + "score": 1.0, + "content": "Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 126, + 306, + 401, + 317 + ], + "spans": [ + { + "bbox": [ + 126, + 306, + 401, + 317 + ], + "score": 1.0, + "content": "2019, December 8-14, 2019, Vancouver, BC, Canada, pages 13–23, 2019. 2", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 324, + 504, + 354 + ], + "lines": [ + { + "bbox": [ + 106, + 324, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 335 + ], + "score": 1.0, + "content": "[36] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 125, + 334, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 125, + 334, + 505, + 346 + ], + "score": 1.0, + "content": "Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 125, + 344, + 338, + 355 + ], + "spans": [ + { + "bbox": [ + 125, + 344, + 338, + 355 + ], + "score": 1.0, + "content": "generation. ArXiv preprint, abs/2002.06353, 2020. 1, 3, 6", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 104, + 362, + 505, + 383 + ], + "lines": [ + { + "bbox": [ + 106, + 362, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 505, + 374 + ], + "score": 1.0, + "content": "[37] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 125, + 372, + 484, + 383 + ], + "spans": [ + { + "bbox": [ + 125, + 372, + 484, + 383 + ], + "score": 1.0, + "content": "empirical study of clip for end to end video clip retrieval. ArXiv preprint, abs/2104.08860, 2021. 3", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 390, + 506, + 431 + ], + "lines": [ + { + "bbox": [ + 106, + 392, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 506, + 402 + ], + "score": 1.0, + "content": "[38] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 126, + 401, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 126, + 401, + 505, + 413 + ], + "score": 1.0, + "content": "End-to-end learning of visual representations from uncurated instructional videos. In 2020 IEEE/CVF", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 126, + 410, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 126, + 410, + 506, + 422 + ], + "score": 1.0, + "content": "Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 125, + 420, + 249, + 432 + ], + "spans": [ + { + "bbox": [ + 125, + 420, + 249, + 432 + ], + "score": 1.0, + "content": "pages 9876–9886. IEEE, 2020. 3", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 106, + 439, + 506, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 439, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 506, + 451 + ], + "score": 1.0, + "content": "[39] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 126, + 449, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 126, + 449, + 506, + 461 + ], + "score": 1.0, + "content": "Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In 2019", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 126, + 458, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 126, + 458, + 506, + 471 + ], + "score": 1.0, + "content": "IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 -", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 125, + 468, + 328, + 481 + ], + "spans": [ + { + "bbox": [ + 125, + 468, + 328, + 481 + ], + "score": 1.0, + "content": "November 2, 2019, pages 2630–2640. IEEE, 2019. 2, 3", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 105, + 487, + 504, + 518 + ], + "lines": [ + { + "bbox": [ + 105, + 487, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 500 + ], + "score": 1.0, + "content": "[40] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 126, + 498, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 126, + 498, + 505, + 510 + ], + "score": 1.0, + "content": "Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 125, + 507, + 350, + 519 + ], + "spans": [ + { + "bbox": [ + 125, + 507, + 350, + 519 + ], + "score": 1.0, + "content": "human feedback. ArXiv preprint, abs/2203.02155, 2022. 2, 6", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 506, + 567 + ], + "lines": [ + { + "bbox": [ + 106, + 526, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 506, + 538 + ], + "score": 1.0, + "content": "[41] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 126, + 536, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 126, + 536, + 505, + 547 + ], + "score": 1.0, + "content": "of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 126, + 546, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 126, + 546, + 505, + 558 + ], + "score": 1.0, + "content": "Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA, 2002. Association for Computational", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 125, + 556, + 180, + 568 + ], + "spans": [ + { + "bbox": [ + 125, + 556, + 180, + 568 + ], + "score": 1.0, + "content": "Linguistics. 6", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 106, + 574, + 506, + 615 + ], + "lines": [ + { + "bbox": [ + 106, + 574, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 506, + 586 + ], + "score": 1.0, + "content": "[42] Mandela Patrick, Po-Yao Huang, Yuki Markus Asano, Florian Metze, Alexander G. Hauptmann, João F.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 125, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 125, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "Henriques, and Andrea Vedaldi. Support-set bottlenecks for video-text representation learning. In 9th", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 125, + 594, + 507, + 607 + ], + "spans": [ + { + "bbox": [ + 125, + 594, + 507, + 607 + ], + "score": 1.0, + "content": "International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 126, + 604, + 221, + 616 + ], + "spans": [ + { + "bbox": [ + 126, + 604, + 221, + 616 + ], + "score": 1.0, + "content": "OpenReview.net, 2021. 3", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 105, + 623, + 506, + 654 + ], + "lines": [ + { + "bbox": [ + 106, + 623, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 506, + 635 + ], + "score": 1.0, + "content": "[43] Baolin Peng, Chenguang Zhu, Chunyuan Li, Xiujun Li, Jinchao Li, Michael Zeng, and Jianfeng Gao. Few-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 126, + 633, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 126, + 633, + 505, + 644 + ], + "score": 1.0, + "content": "shot natural language generation for task-oriented dialog. In Findings of the Association for Computational", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 126, + 643, + 501, + 655 + ], + "spans": [ + { + "bbox": [ + 126, + 643, + 501, + 655 + ], + "score": 1.0, + "content": "Linguistics: EMNLP 2020, pages 172–182, Online, 2020. Association for Computational Linguistics. 1", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 107, + 661, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 661, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 506, + 673 + ], + "score": 1.0, + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 125, + 669, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 125, + 669, + 506, + 685 + ], + "score": 1.0, + "content": "Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 125, + 680, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 125, + 680, + 506, + 694 + ], + "score": 1.0, + "content": "transferable visual models from natural language supervision. In Marina Meila and Tong Zhang, editors,", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 125, + 691, + 507, + 703 + ], + "spans": [ + { + "bbox": [ + 125, + 691, + 507, + 703 + ], + "score": 1.0, + "content": "Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021,", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 126, + 701, + 507, + 713 + ], + "spans": [ + { + "bbox": [ + 126, + 701, + 507, + 713 + ], + "score": 1.0, + "content": "Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 8748–8763. PMLR, 2021.", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 124, + 711, + 153, + 723 + ], + "spans": [ + { + "bbox": [ + 124, + 711, + 153, + 723 + ], + "score": 1.0, + "content": "2, 3, 4", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 49.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 73, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "score": 1.0, + "content": "[30] Chin-Yew Lin and Franz Josef Och. Automatic evaluation of machine translation quality using longest", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 125, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "common subsequence and skip-bigram statistics. In Proceedings of the 42nd Annual Meeting of the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 125, + 93, + 475, + 105 + ], + "spans": [ + { + "bbox": [ + 125, + 93, + 475, + 105 + ], + "score": 1.0, + "content": "Association for Computational Linguistics (ACL-04), pages 605–612, Barcelona, Spain, 2004. 6", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 73, + 505, + 105 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 503, + 143 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 124 + ], + "score": 1.0, + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 121, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 125, + 121, + 505, + 134 + ], + "score": 1.0, + "content": "and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 132, + 313, + 143 + ], + "spans": [ + { + "bbox": [ + 126, + 132, + 313, + 143 + ], + "score": 1.0, + "content": "computer vision, pages 740–755. Springer, 2014. 6", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 111, + 505, + 143 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 150, + 504, + 181 + ], + "lines": [ + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "score": 1.0, + "content": "[32] Xudong Lin, Gedas Bertasius, Jue Wang, Shih-Fu Chang, Devi Parikh, and Lorenzo Torresani. Vx2text:", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 161, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 126, + 161, + 505, + 171 + ], + "score": 1.0, + "content": "End-to-end learning of video-based text generation from multimodal inputs. In 2021 IEEE/CVF Conference", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 125, + 170, + 454, + 182 + ], + "spans": [ + { + "bbox": [ + 125, + 170, + 454, + 182 + ], + "score": 1.0, + "content": "on Computer Vision and Pattern Recognition (CVPR), pages 7001–7011. IEEE, 2021. 3, 4", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 149, + 506, + 182 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 189, + 505, + 219 + ], + "lines": [ + { + "bbox": [ + 105, + 188, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 506, + 200 + ], + "score": 1.0, + "content": "[33] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 124, + 199, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 124, + 199, + 506, + 210 + ], + "score": 1.0, + "content": "Learning to recognize procedural activities with distant supervision. In Proceedings of the IEEE/CVF", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 126, + 209, + 443, + 220 + ], + "spans": [ + { + "bbox": [ + 126, + 209, + 443, + 220 + ], + "score": 1.0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 13853–13863, 2022. 3", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 188, + 506, + 220 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 227, + 505, + 257 + ], + "lines": [ + { + "bbox": [ + 105, + 226, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 506, + 240 + ], + "score": 1.0, + "content": "[34] Xudong Lin, Simran Tiwari, Shiyuan Huang, Manling Li, Mike Zheng Shou, Heng Ji, and Shih-Fu Chang.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 125, + 237, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 125, + 237, + 505, + 248 + ], + "score": 1.0, + "content": "Towards fast adaptation of pretrained contrastive models for multi-channel video-language retrieval. arXiv", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 125, + 247, + 258, + 258 + ], + "spans": [ + { + "bbox": [ + 125, + 247, + 258, + 258 + ], + "score": 1.0, + "content": "preprint arXiv:2206.02082, 2022. 1", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 226, + 506, + 258 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 265, + 505, + 316 + ], + "lines": [ + { + "bbox": [ + 106, + 266, + 504, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 504, + 277 + ], + "score": 1.0, + "content": "[35] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 125, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 125, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "representations for vision-and-language tasks. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 125, + 285, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 125, + 285, + 505, + 297 + ], + "score": 1.0, + "content": "Florence d’Alché-Buc, Emily B. Fox, and Roman Garnett, editors, Advances in Neural Information", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 126, + 296, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 126, + 296, + 505, + 307 + ], + "score": 1.0, + "content": "Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 126, + 306, + 401, + 317 + ], + "spans": [ + { + "bbox": [ + 126, + 306, + 401, + 317 + ], + "score": 1.0, + "content": "2019, December 8-14, 2019, Vancouver, BC, Canada, pages 13–23, 2019. 2", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17, + "bbox_fs": [ + 106, + 266, + 506, + 317 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 324, + 504, + 354 + ], + "lines": [ + { + "bbox": [ + 106, + 324, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 335 + ], + "score": 1.0, + "content": "[36] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 125, + 334, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 125, + 334, + 505, + 346 + ], + "score": 1.0, + "content": "Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 125, + 344, + 338, + 355 + ], + "spans": [ + { + "bbox": [ + 125, + 344, + 338, + 355 + ], + "score": 1.0, + "content": "generation. ArXiv preprint, abs/2002.06353, 2020. 1, 3, 6", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 106, + 324, + 505, + 355 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 362, + 505, + 383 + ], + "lines": [ + { + "bbox": [ + 106, + 362, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 505, + 374 + ], + "score": 1.0, + "content": "[37] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 125, + 372, + 484, + 383 + ], + "spans": [ + { + "bbox": [ + 125, + 372, + 484, + 383 + ], + "score": 1.0, + "content": "empirical study of clip for end to end video clip retrieval. ArXiv preprint, abs/2104.08860, 2021. 3", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 106, + 362, + 505, + 383 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 390, + 506, + 431 + ], + "lines": [ + { + "bbox": [ + 106, + 392, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 506, + 402 + ], + "score": 1.0, + "content": "[38] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 126, + 401, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 126, + 401, + 505, + 413 + ], + "score": 1.0, + "content": "End-to-end learning of visual representations from uncurated instructional videos. In 2020 IEEE/CVF", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 126, + 410, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 126, + 410, + 506, + 422 + ], + "score": 1.0, + "content": "Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 125, + 420, + 249, + 432 + ], + "spans": [ + { + "bbox": [ + 125, + 420, + 249, + 432 + ], + "score": 1.0, + "content": "pages 9876–9886. IEEE, 2020. 3", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5, + "bbox_fs": [ + 106, + 392, + 506, + 432 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 439, + 506, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 439, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 506, + 451 + ], + "score": 1.0, + "content": "[39] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 126, + 449, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 126, + 449, + 506, + 461 + ], + "score": 1.0, + "content": "Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In 2019", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 126, + 458, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 126, + 458, + 506, + 471 + ], + "score": 1.0, + "content": "IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 -", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 125, + 468, + 328, + 481 + ], + "spans": [ + { + "bbox": [ + 125, + 468, + 328, + 481 + ], + "score": 1.0, + "content": "November 2, 2019, pages 2630–2640. IEEE, 2019. 2, 3", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 439, + 506, + 481 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 487, + 504, + 518 + ], + "lines": [ + { + "bbox": [ + 105, + 487, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 500 + ], + "score": 1.0, + "content": "[40] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 126, + 498, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 126, + 498, + 505, + 510 + ], + "score": 1.0, + "content": "Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 125, + 507, + 350, + 519 + ], + "spans": [ + { + "bbox": [ + 125, + 507, + 350, + 519 + ], + "score": 1.0, + "content": "human feedback. ArXiv preprint, abs/2203.02155, 2022. 2, 6", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 487, + 506, + 519 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 525, + 506, + 567 + ], + "lines": [ + { + "bbox": [ + 106, + 526, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 506, + 538 + ], + "score": 1.0, + "content": "[41] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 126, + 536, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 126, + 536, + 505, + 547 + ], + "score": 1.0, + "content": "of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 126, + 546, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 126, + 546, + 505, + 558 + ], + "score": 1.0, + "content": "Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA, 2002. Association for Computational", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 125, + 556, + 180, + 568 + ], + "spans": [ + { + "bbox": [ + 125, + 556, + 180, + 568 + ], + "score": 1.0, + "content": "Linguistics. 6", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5, + "bbox_fs": [ + 106, + 526, + 506, + 568 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 574, + 506, + 615 + ], + "lines": [ + { + "bbox": [ + 106, + 574, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 506, + 586 + ], + "score": 1.0, + "content": "[42] Mandela Patrick, Po-Yao Huang, Yuki Markus Asano, Florian Metze, Alexander G. Hauptmann, João F.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 125, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 125, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "Henriques, and Andrea Vedaldi. Support-set bottlenecks for video-text representation learning. In 9th", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 125, + 594, + 507, + 607 + ], + "spans": [ + { + "bbox": [ + 125, + 594, + 507, + 607 + ], + "score": 1.0, + "content": "International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 126, + 604, + 221, + 616 + ], + "spans": [ + { + "bbox": [ + 126, + 604, + 221, + 616 + ], + "score": 1.0, + "content": "OpenReview.net, 2021. 3", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5, + "bbox_fs": [ + 106, + 574, + 507, + 616 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 623, + 506, + 654 + ], + "lines": [ + { + "bbox": [ + 106, + 623, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 506, + 635 + ], + "score": 1.0, + "content": "[43] Baolin Peng, Chenguang Zhu, Chunyuan Li, Xiujun Li, Jinchao Li, Michael Zeng, and Jianfeng Gao. Few-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 126, + 633, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 126, + 633, + 505, + 644 + ], + "score": 1.0, + "content": "shot natural language generation for task-oriented dialog. In Findings of the Association for Computational", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 126, + 643, + 501, + 655 + ], + "spans": [ + { + "bbox": [ + 126, + 643, + 501, + 655 + ], + "score": 1.0, + "content": "Linguistics: EMNLP 2020, pages 172–182, Online, 2020. Association for Computational Linguistics. 1", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 661, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 506, + 673 + ], + "score": 1.0, + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 125, + 669, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 125, + 669, + 506, + 685 + ], + "score": 1.0, + "content": "Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 125, + 680, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 125, + 680, + 506, + 694 + ], + "score": 1.0, + "content": "transferable visual models from natural language supervision. In Marina Meila and Tong Zhang, editors,", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 125, + 691, + 507, + 703 + ], + "spans": [ + { + "bbox": [ + 125, + 691, + 507, + 703 + ], + "score": 1.0, + "content": "Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021,", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 126, + 701, + 507, + 713 + ], + "spans": [ + { + "bbox": [ + 126, + 701, + 507, + 713 + ], + "score": 1.0, + "content": "Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 8748–8763. PMLR, 2021.", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 124, + 711, + 153, + 723 + ], + "spans": [ + { + "bbox": [ + 124, + 711, + 153, + 723 + ], + "score": 1.0, + "content": "2, 3, 4", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 45, + "bbox_fs": [ + 106, + 623, + 506, + 655 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 661, + 506, + 722 + ], + "lines": [], + "index": 49.5, + "bbox_fs": [ + 105, + 661, + 507, + 723 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 73, + 504, + 104 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 125, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 125, + 93, + 274, + 104 + ], + "spans": [ + { + "bbox": [ + 125, + 93, + 274, + 104 + ], + "score": 1.0, + "content": "ArXiv preprint, abs/1910.10683, 2019. 1", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 105, + 112, + 504, + 143 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "score": 1.0, + "content": "[46] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 122, + 506, + 134 + ], + "spans": [ + { + "bbox": [ + 126, + 122, + 506, + 134 + ], + "score": 1.0, + "content": "Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 132, + 336, + 144 + ], + "spans": [ + { + "bbox": [ + 126, + 132, + 336, + 144 + ], + "score": 1.0, + "content": "Journal of Machine Learning Research, 21:1–67, 2020. 3", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 151, + 505, + 192 + ], + "lines": [ + { + "bbox": [ + 106, + 152, + 506, + 164 + ], + "spans": [ + { + "bbox": [ + 106, + 152, + 506, + 164 + ], + "score": 1.0, + "content": "[47] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using Siamese BERT-networks.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 162, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 126, + 162, + 506, + 174 + ], + "score": 1.0, + "content": "In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 126, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 126, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982–3992,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 126, + 181, + 405, + 194 + ], + "spans": [ + { + "bbox": [ + 126, + 181, + 405, + 194 + ], + "score": 1.0, + "content": "Hong Kong, China, 2019. Association for Computational Linguistics. 4, 6, 8", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 200, + 504, + 241 + ], + "lines": [ + { + "bbox": [ + 106, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "[48] Timo Schick and Hinrich Schütze. Exploiting cloze-questions for few-shot text classification and natural", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 126, + 211, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 126, + 211, + 505, + 223 + ], + "score": 1.0, + "content": "language inference. In Proceedings of the 16th Conference of the European Chapter of the Association for", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 126, + 221, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 126, + 221, + 505, + 232 + ], + "score": 1.0, + "content": "Computational Linguistics: Main Volume, pages 255–269, Online, 2021. Association for Computational", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 126, + 231, + 179, + 242 + ], + "spans": [ + { + "bbox": [ + 126, + 231, + 179, + 242 + ], + "score": 1.0, + "content": "Linguistics. 2", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 105, + 249, + 505, + 271 + ], + "lines": [ + { + "bbox": [ + 106, + 249, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 505, + 263 + ], + "score": 1.0, + "content": "[49] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 125, + 259, + 396, + 271 + ], + "spans": [ + { + "bbox": [ + 125, + 259, + 396, + 271 + ], + "score": 1.0, + "content": "for multimodal video captioning. ArXiv preprint, abs/2201.08264, 2022. 1", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 279, + 504, + 310 + ], + "lines": [ + { + "bbox": [ + 106, + 279, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 506, + 291 + ], + "score": 1.0, + "content": "[50] Yixuan Su, Tian Lan, Yahui Liu, Fangyu Liu, Dani Yogatama, Yan Wang, Lingpeng Kong, and Nigel Collier.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 125, + 289, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 125, + 289, + 506, + 301 + ], + "score": 1.0, + "content": "Language models can see: Plugging visual controls in text generation. ArXiv preprint, abs/2205.02655,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 125, + 299, + 158, + 310 + ], + "spans": [ + { + "bbox": [ + 125, + 299, + 158, + 310 + ], + "score": 1.0, + "content": "2022. 3", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 318, + 506, + 359 + ], + "lines": [ + { + "bbox": [ + 106, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "[51] Derek Tam, Rakesh R. Menon, Mohit Bansal, Shashank Srivastava, and Colin Raffel. Improving and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 126, + 329, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 126, + 329, + 506, + 340 + ], + "score": 1.0, + "content": "simplifying pattern exploiting training. In Proceedings of the 2021 Conference on Empirical Methods", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 126, + 338, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 126, + 338, + 506, + 350 + ], + "score": 1.0, + "content": "in Natural Language Processing, pages 4980–4991, Online and Punta Cana, Dominican Republic, 2021.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 126, + 348, + 292, + 360 + ], + "spans": [ + { + "bbox": [ + 126, + 348, + 292, + 360 + ], + "score": 1.0, + "content": "Association for Computational Linguistics. 2", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 367, + 506, + 408 + ], + "lines": [ + { + "bbox": [ + 106, + 367, + 507, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 507, + 380 + ], + "score": 1.0, + "content": "[52] Hao Tan and Mohit Bansal. LXMERT: Learning cross-modality encoder representations from transformers.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 126, + 378, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 126, + 378, + 505, + 389 + ], + "score": 1.0, + "content": "In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 126, + 387, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 126, + 387, + 506, + 399 + ], + "score": 1.0, + "content": "International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5100–5111,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 126, + 397, + 387, + 410 + ], + "spans": [ + { + "bbox": [ + 126, + 397, + 387, + 410 + ], + "score": 1.0, + "content": "Hong Kong, China, 2019. Association for Computational Linguistics. 2", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 416, + 504, + 447 + ], + "lines": [ + { + "bbox": [ + 105, + 416, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 428 + ], + "score": 1.0, + "content": "[53] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill. Multimodal", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 125, + 426, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 125, + 426, + 506, + 439 + ], + "score": 1.0, + "content": "few-shot learning with frozen language models. Advances in Neural Information Processing Systems,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 125, + 437, + 205, + 447 + ], + "spans": [ + { + "bbox": [ + 125, + 437, + 205, + 447 + ], + "score": 1.0, + "content": "34:200–212, 2021. 3", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 456, + 504, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 455, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 506, + 468 + ], + "score": 1.0, + "content": "[54] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 125, + 466, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 125, + 466, + 506, + 478 + ], + "score": 1.0, + "content": "evaluation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 126, + 476, + 399, + 488 + ], + "spans": [ + { + "bbox": [ + 126, + 476, + 399, + 488 + ], + "score": 1.0, + "content": "USA, June 7-12, 2015, pages 4566–4575. IEEE Computer Society, 2015. 6", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 105, + 495, + 504, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 495, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 506, + 507 + ], + "score": 1.0, + "content": "[55] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 125, + 505, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 125, + 505, + 506, + 517 + ], + "score": 1.0, + "content": "Zhou, and Hongxia Yang. Unifying architectures, tasks, and modalities through a simple sequence-to-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 126, + 516, + 387, + 527 + ], + "spans": [ + { + "bbox": [ + 126, + 516, + 387, + 527 + ], + "score": 1.0, + "content": "sequence learning framework. ArXiv preprint, abs/2202.03052, 2022. 3", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 105, + 534, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 533, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 547 + ], + "score": 1.0, + "content": "[56] Qiang Wang, Yanhao Zhang, Yun Zheng, Pan Pan, and Xian-Sheng Hua. Disentangled representation", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 124, + 544, + 394, + 556 + ], + "spans": [ + { + "bbox": [ + 124, + 544, + 394, + 556 + ], + "score": 1.0, + "content": "learning for text-video retrieval. ArXiv preprint, abs/2203.07111, 2022. 9", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 106, + 563, + 506, + 604 + ], + "lines": [ + { + "bbox": [ + 106, + 563, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 506, + 575 + ], + "score": 1.0, + "content": "[57] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 126, + 574, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 126, + 574, + 506, + 586 + ], + "score": 1.0, + "content": "scale, high-quality multilingual dataset for video-and-language research. In 2019 IEEE/CVF International", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 125, + 582, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 125, + 582, + 506, + 597 + ], + "score": 1.0, + "content": "Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 125, + 592, + 236, + 605 + ], + "spans": [ + { + "bbox": [ + 125, + 592, + 236, + 605 + ], + "score": 1.0, + "content": "4580–4590. IEEE, 2019. 1, 6", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 104, + 613, + 504, + 634 + ], + "lines": [ + { + "bbox": [ + 106, + 613, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 505, + 624 + ], + "score": 1.0, + "content": "[58] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 126, + 623, + 463, + 634 + ], + "spans": [ + { + "bbox": [ + 126, + 623, + 463, + 634 + ], + "score": 1.0, + "content": "language model pretraining with weak supervision. ArXiv preprint, abs/2108.10904, 2021. 2", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 105, + 642, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 105, + 642, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 505, + 653 + ], + "score": 1.0, + "content": "[59] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 125, + 652, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 125, + 652, + 506, + 664 + ], + "score": 1.0, + "content": "Dai, and Quoc V Le. Finetuned language models are zero-shot learners. ArXiv preprint, abs/2109.01652,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 124, + 660, + 159, + 673 + ], + "spans": [ + { + "bbox": [ + 124, + 660, + 159, + 673 + ], + "score": 1.0, + "content": "2021. 5", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 107, + 681, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 681, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 505, + 693 + ], + "score": 1.0, + "content": "[60] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 125, + 691, + 506, + 704 + ], + "spans": [ + { + "bbox": [ + 125, + 691, + 506, + 704 + ], + "score": 1.0, + "content": "question answering via gradually refined attention over appearance and motion. In Proceedings of the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 124, + 700, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 124, + 700, + 505, + 714 + ], + "score": 1.0, + "content": "2017 ACM on Multimedia Conference, MM 2017, Mountain View, CA, USA, October 23-27, 2017, pages", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 126, + 712, + 201, + 722 + ], + "spans": [ + { + "bbox": [ + 126, + 712, + 201, + 722 + ], + "score": 1.0, + "content": "1645–1653, 2017. 1", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 73, + 504, + 104 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 125, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 125, + 93, + 274, + 104 + ], + "spans": [ + { + "bbox": [ + 125, + 93, + 274, + 104 + ], + "score": 1.0, + "content": "ArXiv preprint, abs/1910.10683, 2019. 1", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 72, + 506, + 104 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 112, + 504, + 143 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "score": 1.0, + "content": "[46] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 122, + 506, + 134 + ], + "spans": [ + { + "bbox": [ + 126, + 122, + 506, + 134 + ], + "score": 1.0, + "content": "Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 132, + 336, + 144 + ], + "spans": [ + { + "bbox": [ + 126, + 132, + 336, + 144 + ], + "score": 1.0, + "content": "Journal of Machine Learning Research, 21:1–67, 2020. 3", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 112, + 506, + 144 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 151, + 505, + 192 + ], + "lines": [ + { + "bbox": [ + 106, + 152, + 506, + 164 + ], + "spans": [ + { + "bbox": [ + 106, + 152, + 506, + 164 + ], + "score": 1.0, + "content": "[47] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using Siamese BERT-networks.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 162, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 126, + 162, + 506, + 174 + ], + "score": 1.0, + "content": "In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 126, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 126, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982–3992,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 126, + 181, + 405, + 194 + ], + "spans": [ + { + "bbox": [ + 126, + 181, + 405, + 194 + ], + "score": 1.0, + "content": "Hong Kong, China, 2019. Association for Computational Linguistics. 4, 6, 8", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 152, + 506, + 194 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 200, + 504, + 241 + ], + "lines": [ + { + "bbox": [ + 106, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "[48] Timo Schick and Hinrich Schütze. Exploiting cloze-questions for few-shot text classification and natural", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 126, + 211, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 126, + 211, + 505, + 223 + ], + "score": 1.0, + "content": "language inference. In Proceedings of the 16th Conference of the European Chapter of the Association for", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 126, + 221, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 126, + 221, + 505, + 232 + ], + "score": 1.0, + "content": "Computational Linguistics: Main Volume, pages 255–269, Online, 2021. Association for Computational", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 126, + 231, + 179, + 242 + ], + "spans": [ + { + "bbox": [ + 126, + 231, + 179, + 242 + ], + "score": 1.0, + "content": "Linguistics. 2", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5, + "bbox_fs": [ + 106, + 201, + 505, + 242 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 249, + 505, + 271 + ], + "lines": [ + { + "bbox": [ + 106, + 249, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 505, + 263 + ], + "score": 1.0, + "content": "[49] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 125, + 259, + 396, + 271 + ], + "spans": [ + { + "bbox": [ + 125, + 259, + 396, + 271 + ], + "score": 1.0, + "content": "for multimodal video captioning. ArXiv preprint, abs/2201.08264, 2022. 1", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 249, + 505, + 271 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 279, + 504, + 310 + ], + "lines": [ + { + "bbox": [ + 106, + 279, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 506, + 291 + ], + "score": 1.0, + "content": "[50] Yixuan Su, Tian Lan, Yahui Liu, Fangyu Liu, Dani Yogatama, Yan Wang, Lingpeng Kong, and Nigel Collier.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 125, + 289, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 125, + 289, + 506, + 301 + ], + "score": 1.0, + "content": "Language models can see: Plugging visual controls in text generation. ArXiv preprint, abs/2205.02655,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 125, + 299, + 158, + 310 + ], + "spans": [ + { + "bbox": [ + 125, + 299, + 158, + 310 + ], + "score": 1.0, + "content": "2022. 3", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 106, + 279, + 506, + 310 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 318, + 506, + 359 + ], + "lines": [ + { + "bbox": [ + 106, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "[51] Derek Tam, Rakesh R. Menon, Mohit Bansal, Shashank Srivastava, and Colin Raffel. Improving and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 126, + 329, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 126, + 329, + 506, + 340 + ], + "score": 1.0, + "content": "simplifying pattern exploiting training. In Proceedings of the 2021 Conference on Empirical Methods", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 126, + 338, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 126, + 338, + 506, + 350 + ], + "score": 1.0, + "content": "in Natural Language Processing, pages 4980–4991, Online and Punta Cana, Dominican Republic, 2021.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 126, + 348, + 292, + 360 + ], + "spans": [ + { + "bbox": [ + 126, + 348, + 292, + 360 + ], + "score": 1.0, + "content": "Association for Computational Linguistics. 2", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5, + "bbox_fs": [ + 106, + 318, + 506, + 360 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 367, + 506, + 408 + ], + "lines": [ + { + "bbox": [ + 106, + 367, + 507, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 507, + 380 + ], + "score": 1.0, + "content": "[52] Hao Tan and Mohit Bansal. LXMERT: Learning cross-modality encoder representations from transformers.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 126, + 378, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 126, + 378, + 505, + 389 + ], + "score": 1.0, + "content": "In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 126, + 387, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 126, + 387, + 506, + 399 + ], + "score": 1.0, + "content": "International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5100–5111,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 126, + 397, + 387, + 410 + ], + "spans": [ + { + "bbox": [ + 126, + 397, + 387, + 410 + ], + "score": 1.0, + "content": "Hong Kong, China, 2019. Association for Computational Linguistics. 2", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 106, + 367, + 507, + 410 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 416, + 504, + 447 + ], + "lines": [ + { + "bbox": [ + 105, + 416, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 428 + ], + "score": 1.0, + "content": "[53] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill. Multimodal", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 125, + 426, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 125, + 426, + 506, + 439 + ], + "score": 1.0, + "content": "few-shot learning with frozen language models. Advances in Neural Information Processing Systems,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 125, + 437, + 205, + 447 + ], + "spans": [ + { + "bbox": [ + 125, + 437, + 205, + 447 + ], + "score": 1.0, + "content": "34:200–212, 2021. 3", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 416, + 506, + 447 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 456, + 504, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 455, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 506, + 468 + ], + "score": 1.0, + "content": "[54] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 125, + 466, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 125, + 466, + 506, + 478 + ], + "score": 1.0, + "content": "evaluation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 126, + 476, + 399, + 488 + ], + "spans": [ + { + "bbox": [ + 126, + 476, + 399, + 488 + ], + "score": 1.0, + "content": "USA, June 7-12, 2015, pages 4566–4575. IEEE Computer Society, 2015. 6", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 455, + 506, + 488 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 495, + 504, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 495, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 506, + 507 + ], + "score": 1.0, + "content": "[55] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 125, + 505, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 125, + 505, + 506, + 517 + ], + "score": 1.0, + "content": "Zhou, and Hongxia Yang. Unifying architectures, tasks, and modalities through a simple sequence-to-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 126, + 516, + 387, + 527 + ], + "spans": [ + { + "bbox": [ + 126, + 516, + 387, + 527 + ], + "score": 1.0, + "content": "sequence learning framework. ArXiv preprint, abs/2202.03052, 2022. 3", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 495, + 506, + 527 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 534, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 533, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 547 + ], + "score": 1.0, + "content": "[56] Qiang Wang, Yanhao Zhang, Yun Zheng, Pan Pan, and Xian-Sheng Hua. Disentangled representation", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 124, + 544, + 394, + 556 + ], + "spans": [ + { + "bbox": [ + 124, + 544, + 394, + 556 + ], + "score": 1.0, + "content": "learning for text-video retrieval. ArXiv preprint, abs/2203.07111, 2022. 9", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 106, + 533, + 505, + 556 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 563, + 506, + 604 + ], + "lines": [ + { + "bbox": [ + 106, + 563, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 506, + 575 + ], + "score": 1.0, + "content": "[57] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. Vatex: A large-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 126, + 574, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 126, + 574, + 506, + 586 + ], + "score": 1.0, + "content": "scale, high-quality multilingual dataset for video-and-language research. In 2019 IEEE/CVF International", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 125, + 582, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 125, + 582, + 506, + 597 + ], + "score": 1.0, + "content": "Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 125, + 592, + 236, + 605 + ], + "spans": [ + { + "bbox": [ + 125, + 592, + 236, + 605 + ], + "score": 1.0, + "content": "4580–4590. IEEE, 2019. 1, 6", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5, + "bbox_fs": [ + 106, + 563, + 506, + 605 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 613, + 504, + 634 + ], + "lines": [ + { + "bbox": [ + 106, + 613, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 505, + 624 + ], + "score": 1.0, + "content": "[58] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 126, + 623, + 463, + 634 + ], + "spans": [ + { + "bbox": [ + 126, + 623, + 463, + 634 + ], + "score": 1.0, + "content": "language model pretraining with weak supervision. ArXiv preprint, abs/2108.10904, 2021. 2", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 106, + 613, + 505, + 634 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 642, + 505, + 673 + ], + "lines": [ + { + "bbox": [ + 105, + 642, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 505, + 653 + ], + "score": 1.0, + "content": "[59] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 125, + 652, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 125, + 652, + 506, + 664 + ], + "score": 1.0, + "content": "Dai, and Quoc V Le. Finetuned language models are zero-shot learners. ArXiv preprint, abs/2109.01652,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 124, + 660, + 159, + 673 + ], + "spans": [ + { + "bbox": [ + 124, + 660, + 159, + 673 + ], + "score": 1.0, + "content": "2021. 5", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 642, + 506, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 681, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 681, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 505, + 693 + ], + "score": 1.0, + "content": "[60] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 125, + 691, + 506, + 704 + ], + "spans": [ + { + "bbox": [ + 125, + 691, + 506, + 704 + ], + "score": 1.0, + "content": "question answering via gradually refined attention over appearance and motion. In Proceedings of the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 124, + 700, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 124, + 700, + 505, + 714 + ], + "score": 1.0, + "content": "2017 ACM on Multimedia Conference, MM 2017, Mountain View, CA, USA, October 23-27, 2017, pages", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 126, + 712, + 201, + 722 + ], + "spans": [ + { + "bbox": [ + 126, + 712, + 201, + 722 + ], + "score": 1.0, + "content": "1645–1653, 2017. 1", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48.5, + "bbox_fs": [ + 106, + 681, + 506, + 722 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 28, + 507, + 697 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 507, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 507, + 86 + ], + "score": 1.0, + "content": "[61] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 82, + 507, + 96 + ], + "spans": [ + { + "bbox": [ + 125, + 82, + 507, + 96 + ], + "score": 1.0, + "content": "Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pre-training for zero-shot video-text", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 126, + 94, + 505, + 105 + ], + "score": 1.0, + "content": "understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 102, + 507, + 117 + ], + "spans": [ + { + "bbox": [ + 124, + 102, + 507, + 117 + ], + "score": 1.0, + "content": "Processing, pages 6787–6800, Online and Punta Cana, Dominican Republic, 2021. Association for", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 124, + 111, + 237, + 126 + ], + "spans": [ + { + "bbox": [ + 124, + 111, + 237, + 126 + ], + "score": 1.0, + "content": "Computational Linguistics. 1", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 128, + 507, + 145 + ], + "spans": [ + { + "bbox": [ + 104, + 128, + 507, + 145 + ], + "score": 1.0, + "content": "[62] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 141, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 126, + 141, + 505, + 153 + ], + "score": 1.0, + "content": "video and language. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 125, + 150, + 467, + 164 + ], + "spans": [ + { + "bbox": [ + 125, + 150, + 467, + 164 + ], + "score": 1.0, + "content": "Las Vegas, NV, USA, June 27-30, 2016, pages 5288–5296. IEEE Computer Society, 2016. 1, 6", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 168, + 507, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 507, + 182 + ], + "score": 1.0, + "content": "[63] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Yumao Lu, Zicheng Liu, and Lijuan Wang. An", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 125, + 178, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 125, + 178, + 505, + 192 + ], + "score": 1.0, + "content": "empirical study of gpt-3 for few-shot knowledge-based vqa. ArXiv preprint, abs/2109.05014, 2021. 1, 3, 4", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 197, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 197, + 506, + 209 + ], + "score": 1.0, + "content": "[64] Ziyi Yang, Yuwei Fang, Chenguang Zhu, Reid Pryzant, Dongdong Chen, Yu Shi, Yichong Xu, Yao Qian,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 124, + 205, + 507, + 220 + ], + "spans": [ + { + "bbox": [ + 124, + 205, + 507, + 220 + ], + "score": 1.0, + "content": "Mei Gao, Yi-Ling Chen, et al. i-code: An integrative and composable multimodal learning framework.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 124, + 216, + 276, + 229 + ], + "spans": [ + { + "bbox": [ + 124, + 216, + 276, + 229 + ], + "score": 1.0, + "content": "ArXiv preprint, abs/2205.01818, 2022. 1", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 235, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 505, + 246 + ], + "score": 1.0, + "content": "[65] Fei Yu, Jiji Tang, Weichong Yin, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Ernie-vil: Knowledge", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 125, + 244, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 125, + 244, + 506, + 258 + ], + "score": 1.0, + "content": "enhanced vision-language representations through scene graphs. In Proceedings of the AAAI Conference", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 125, + 255, + 360, + 266 + ], + "spans": [ + { + "bbox": [ + 125, + 255, + 360, + 266 + ], + "score": 1.0, + "content": "on Artificial Intelligence, volume 35, pages 3208–3216, 2021. 2", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 270, + 507, + 287 + ], + "spans": [ + { + "bbox": [ + 104, + 270, + 507, + 287 + ], + "score": 1.0, + "content": "[66] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 124, + 282, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 124, + 282, + 506, + 296 + ], + "score": 1.0, + "content": "Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. ArXiv", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 124, + 292, + 253, + 305 + ], + "spans": [ + { + "bbox": [ + 124, + 292, + 253, + 305 + ], + "score": 1.0, + "content": "preprint, abs/2111.11432, 2021. 2", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "score": 1.0, + "content": "[67] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 126, + 321, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 126, + 321, + 505, + 333 + ], + "score": 1.0, + "content": "Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 124, + 330, + 392, + 343 + ], + "spans": [ + { + "bbox": [ + 124, + 330, + 392, + 343 + ], + "score": 1.0, + "content": "vision and language and sound. ArXiv preprint, abs/2201.02639, 2022. 1", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 347, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 505, + 361 + ], + "score": 1.0, + "content": "[68] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 124, + 356, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 124, + 356, + 505, + 372 + ], + "score": 1.0, + "content": "Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 126, + 369, + 222, + 380 + ], + "spans": [ + { + "bbox": [ + 126, + 369, + 222, + 380 + ], + "score": 1.0, + "content": "Systems, 34, 2021. 1, 2, 8", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 386, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 505, + 398 + ], + "score": 1.0, + "content": "[69] Andy Zeng, Adrian Wong, Stefan Welker, Krzysztof Choromanski, Federico Tombari, Aveek Purohit,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 124, + 394, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 124, + 394, + 506, + 410 + ], + "score": 1.0, + "content": "Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, et al. Socratic models: Composing", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 124, + 405, + 458, + 419 + ], + "spans": [ + { + "bbox": [ + 124, + 405, + 458, + 419 + ], + "score": 1.0, + "content": "zero-shot multimodal reasoning with language. ArXiv preprint, abs/2204.00598, 2022. 3, 4", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 423, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 506, + 437 + ], + "score": 1.0, + "content": "[70] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 125, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 125, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 126, + 444, + 474, + 456 + ], + "spans": [ + { + "bbox": [ + 126, + 444, + 474, + 456 + ], + "score": 1.0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579–5588, 2021. 2", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "[71] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 125, + 471, + 507, + 485 + ], + "spans": [ + { + "bbox": [ + 125, + 471, + 507, + 485 + ], + "score": 1.0, + "content": "Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 124, + 481, + 276, + 495 + ], + "spans": [ + { + "bbox": [ + 124, + 481, + 276, + 495 + ], + "score": 1.0, + "content": "ArXiv preprint, abs/2205.01068, 2022. 1", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 497, + 507, + 514 + ], + "spans": [ + { + "bbox": [ + 104, + 497, + 507, + 514 + ], + "score": 1.0, + "content": "[72] Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. Calibrate before use: Improving", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 125, + 509, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 125, + 509, + 505, + 523 + ], + "score": 1.0, + "content": "few-shot performance of language models. In Marina Meila and Tong Zhang, editors, Proceedings of the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 125, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 125, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 125, + 529, + 460, + 542 + ], + "spans": [ + { + "bbox": [ + 125, + 529, + 460, + 542 + ], + "score": 1.0, + "content": "139 of Proceedings of Machine Learning Research, pages 12697–12706. PMLR, 2021. 1, 6", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "[73] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason J. Corso, and Jianfeng Gao. Unified", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 125, + 557, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 125, + 557, + 506, + 571 + ], + "score": 1.0, + "content": "vision-language pre-training for image captioning and VQA. In The Thirty-Fourth AAAI Conference on", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 125, + 568, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 125, + 568, + 505, + 579 + ], + "score": 1.0, + "content": "Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 126, + 578, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 126, + 578, + 506, + 590 + ], + "score": 1.0, + "content": "Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 125, + 587, + 479, + 601 + ], + "spans": [ + { + "bbox": [ + 125, + 587, + 479, + 601 + ], + "score": 1.0, + "content": "EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 13041–13049. AAAI Press, 2020. 2", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "[74] Luowei Zhou, Chenliang Xu, and Jason J. Corso. Towards automatic learning of procedures from web", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 124, + 614, + 507, + 628 + ], + "spans": [ + { + "bbox": [ + 124, + 614, + 507, + 628 + ], + "score": 1.0, + "content": "instructional videos. In Sheila A. McIlraith and Kilian Q. Weinberger, editors, Proceedings of the Thirty-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 126, + 626, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 126, + 626, + 505, + 637 + ], + "score": 1.0, + "content": "Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 125, + 635, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 125, + 635, + 505, + 647 + ], + "score": 1.0, + "content": "Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 125, + 645, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 125, + 645, + 505, + 657 + ], + "score": 1.0, + "content": "(EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pages 7590–7598. AAAI Press, 2018. 3, 6", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 662, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 506, + 676 + ], + "score": 1.0, + "content": "[75] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Xiaogang Wang, Hongsheng Li, Xiaohua Wang, and Jifeng", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 126, + 673, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 126, + 673, + 505, + 685 + ], + "score": 1.0, + "content": "Dai. Uni-perceiver: Pre-training unified architecture for generic perception for zero-shot and few-shot", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 125, + 682, + 300, + 696 + ], + "spans": [ + { + "bbox": [ + 125, + 682, + 300, + 696 + ], + "score": 1.0, + "content": "tasks. ArXiv preprint, abs/2112.01522, 2021. 3", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 25 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 28, + 507, + 697 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 507, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 507, + 86 + ], + "score": 1.0, + "content": "[61] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 82, + 507, + 96 + ], + "spans": [ + { + "bbox": [ + 125, + 82, + 507, + 96 + ], + "score": 1.0, + "content": "Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pre-training for zero-shot video-text", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 126, + 94, + 505, + 105 + ], + "score": 1.0, + "content": "understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 102, + 507, + 117 + ], + "spans": [ + { + "bbox": [ + 124, + 102, + 507, + 117 + ], + "score": 1.0, + "content": "Processing, pages 6787–6800, Online and Punta Cana, Dominican Republic, 2021. Association for", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 124, + 111, + 237, + 126 + ], + "spans": [ + { + "bbox": [ + 124, + 111, + 237, + 126 + ], + "score": 1.0, + "content": "Computational Linguistics. 1", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 128, + 507, + 145 + ], + "spans": [ + { + "bbox": [ + 104, + 128, + 507, + 145 + ], + "score": 1.0, + "content": "[62] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 141, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 126, + 141, + 505, + 153 + ], + "score": 1.0, + "content": "video and language. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 125, + 150, + 467, + 164 + ], + "spans": [ + { + "bbox": [ + 125, + 150, + 467, + 164 + ], + "score": 1.0, + "content": "Las Vegas, NV, USA, June 27-30, 2016, pages 5288–5296. IEEE Computer Society, 2016. 1, 6", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 168, + 507, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 507, + 182 + ], + "score": 1.0, + "content": "[63] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Yumao Lu, Zicheng Liu, and Lijuan Wang. An", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 125, + 178, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 125, + 178, + 505, + 192 + ], + "score": 1.0, + "content": "empirical study of gpt-3 for few-shot knowledge-based vqa. ArXiv preprint, abs/2109.05014, 2021. 1, 3, 4", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 197, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 197, + 506, + 209 + ], + "score": 1.0, + "content": "[64] Ziyi Yang, Yuwei Fang, Chenguang Zhu, Reid Pryzant, Dongdong Chen, Yu Shi, Yichong Xu, Yao Qian,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 124, + 205, + 507, + 220 + ], + "spans": [ + { + "bbox": [ + 124, + 205, + 507, + 220 + ], + "score": 1.0, + "content": "Mei Gao, Yi-Ling Chen, et al. i-code: An integrative and composable multimodal learning framework.", + "type": "text" + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 124, + 216, + 276, + 229 + ], + "spans": [ + { + "bbox": [ + 124, + 216, + 276, + 229 + ], + "score": 1.0, + "content": "ArXiv preprint, abs/2205.01818, 2022. 1", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 235, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 505, + 246 + ], + "score": 1.0, + "content": "[65] Fei Yu, Jiji Tang, Weichong Yin, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Ernie-vil: Knowledge", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 125, + 244, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 125, + 244, + 506, + 258 + ], + "score": 1.0, + "content": "enhanced vision-language representations through scene graphs. In Proceedings of the AAAI Conference", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 125, + 255, + 360, + 266 + ], + "spans": [ + { + "bbox": [ + 125, + 255, + 360, + 266 + ], + "score": 1.0, + "content": "on Artificial Intelligence, volume 35, pages 3208–3216, 2021. 2", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 270, + 507, + 287 + ], + "spans": [ + { + "bbox": [ + 104, + 270, + 507, + 287 + ], + "score": 1.0, + "content": "[66] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 124, + 282, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 124, + 282, + 506, + 296 + ], + "score": 1.0, + "content": "Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. ArXiv", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 124, + 292, + 253, + 305 + ], + "spans": [ + { + "bbox": [ + 124, + 292, + 253, + 305 + ], + "score": 1.0, + "content": "preprint, abs/2111.11432, 2021. 2", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "score": 1.0, + "content": "[67] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 126, + 321, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 126, + 321, + 505, + 333 + ], + "score": 1.0, + "content": "Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 124, + 330, + 392, + 343 + ], + "spans": [ + { + "bbox": [ + 124, + 330, + 392, + 343 + ], + "score": 1.0, + "content": "vision and language and sound. ArXiv preprint, abs/2201.02639, 2022. 1", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 347, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 505, + 361 + ], + "score": 1.0, + "content": "[68] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 124, + 356, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 124, + 356, + 505, + 372 + ], + "score": 1.0, + "content": "Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 126, + 369, + 222, + 380 + ], + "spans": [ + { + "bbox": [ + 126, + 369, + 222, + 380 + ], + "score": 1.0, + "content": "Systems, 34, 2021. 1, 2, 8", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 386, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 505, + 398 + ], + "score": 1.0, + "content": "[69] Andy Zeng, Adrian Wong, Stefan Welker, Krzysztof Choromanski, Federico Tombari, Aveek Purohit,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 124, + 394, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 124, + 394, + 506, + 410 + ], + "score": 1.0, + "content": "Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, et al. Socratic models: Composing", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 124, + 405, + 458, + 419 + ], + "spans": [ + { + "bbox": [ + 124, + 405, + 458, + 419 + ], + "score": 1.0, + "content": "zero-shot multimodal reasoning with language. ArXiv preprint, abs/2204.00598, 2022. 3, 4", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 423, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 506, + 437 + ], + "score": 1.0, + "content": "[70] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 125, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 125, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 126, + 444, + 474, + 456 + ], + "spans": [ + { + "bbox": [ + 126, + 444, + 474, + 456 + ], + "score": 1.0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579–5588, 2021. 2", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "[71] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 125, + 471, + 507, + 485 + ], + "spans": [ + { + "bbox": [ + 125, + 471, + 507, + 485 + ], + "score": 1.0, + "content": "Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models.", + "type": "text" + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 124, + 481, + 276, + 495 + ], + "spans": [ + { + "bbox": [ + 124, + 481, + 276, + 495 + ], + "score": 1.0, + "content": "ArXiv preprint, abs/2205.01068, 2022. 1", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 497, + 507, + 514 + ], + "spans": [ + { + "bbox": [ + 104, + 497, + 507, + 514 + ], + "score": 1.0, + "content": "[72] Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. Calibrate before use: Improving", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 125, + 509, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 125, + 509, + 505, + 523 + ], + "score": 1.0, + "content": "few-shot performance of language models. In Marina Meila and Tong Zhang, editors, Proceedings of the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 125, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 125, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 529, + 460, + 542 + ], + "spans": [ + { + "bbox": [ + 125, + 529, + 460, + 542 + ], + "score": 1.0, + "content": "139 of Proceedings of Machine Learning Research, pages 12697–12706. PMLR, 2021. 1, 6", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "[73] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason J. Corso, and Jianfeng Gao. Unified", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 125, + 557, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 125, + 557, + 506, + 571 + ], + "score": 1.0, + "content": "vision-language pre-training for image captioning and VQA. In The Thirty-Fourth AAAI Conference on", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 125, + 568, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 125, + 568, + 505, + 579 + ], + "score": 1.0, + "content": "Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 126, + 578, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 126, + 578, + 506, + 590 + ], + "score": 1.0, + "content": "Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 125, + 587, + 479, + 601 + ], + "spans": [ + { + "bbox": [ + 125, + 587, + 479, + 601 + ], + "score": 1.0, + "content": "EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 13041–13049. AAAI Press, 2020. 2", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "[74] Luowei Zhou, Chenliang Xu, and Jason J. Corso. Towards automatic learning of procedures from web", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 124, + 614, + 507, + 628 + ], + "spans": [ + { + "bbox": [ + 124, + 614, + 507, + 628 + ], + "score": 1.0, + "content": "instructional videos. In Sheila A. McIlraith and Kilian Q. Weinberger, editors, Proceedings of the Thirty-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 126, + 626, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 126, + 626, + 505, + 637 + ], + "score": 1.0, + "content": "Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 125, + 635, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 125, + 635, + 505, + 647 + ], + "score": 1.0, + "content": "Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 125, + 645, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 125, + 645, + 505, + 657 + ], + "score": 1.0, + "content": "(EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pages 7590–7598. AAAI Press, 2018. 3, 6", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 662, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 506, + 676 + ], + "score": 1.0, + "content": "[75] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Xiaogang Wang, Hongsheng Li, Xiaohua Wang, and Jifeng", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 126, + 673, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 126, + 673, + 505, + 685 + ], + "score": 1.0, + "content": "Dai. Uni-perceiver: Pre-training unified architecture for generic perception for zero-shot and few-shot", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 125, + 682, + 300, + 696 + ], + "spans": [ + { + "bbox": [ + 125, + 682, + 300, + 696 + ], + "score": 1.0, + "content": "tasks. ArXiv preprint, abs/2112.01522, 2021. 3", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 25, + "bbox_fs": [ + 104, + 72, + 507, + 696 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file