pith. machine review for the scientific record. sign in

arxiv: 2604.21873 · v1 · submitted 2026-04-23 · 💻 cs.CV

Recognition: unknown

Grounding Video Reasoning in Physical Signals

Alibay Osmanli, Shaogang Gong, Zixu Cheng

Authors on Pith no claims yet

Pith reviewed 2026-05-09 22:25 UTC · model grok-4.3

classification 💻 cs.CV
keywords video reasoningphysical understandinggrounded benchmarkprompt familiesperturbation analysisevent recordsspatial temporal groundingvideo QA
0
0 comments X

The pith

Video Q&A benchmarks must report physically grounded, prompt-aware, and perturbation-aware diagnostics alongside aggregate accuracy.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that video models can answer questions about events like pouring or collisions by relying on textual patterns in training data, without correctly locating those events in time or space. To expose this gap it builds a benchmark that first converts clips from four sources into shared grounded event records, then derives three prompt families and four input conditions from those records. A sympathetic reader would care because standard accuracy numbers hide whether any actual physical understanding is present. If the central claim holds, evaluations of video reasoning will shift from single scores to sets of targeted checks that separate real grounding from shortcut solutions.

Core claim

The central claim on the paper's own terms is that physics prompts remain the strongest regime overall, vstar-like prompts supply the clearest non-physics semantic baseline, neutral restricted prompts act as harder templated controls, spatial grounding is the weakest dimension across settings, prompt-family robustness is selective rather than universal, and gains from perturbations cluster in originally weak cases; these patterns together imply that video Q&A reasoning benchmarks must report physically grounded, prompt-aware, and perturbation-aware diagnostics in addition to aggregate accuracy.

What carries the argument

The shared grounded event record created from each raw video clip, which supplies consistent temporal and spatial targets plus family-specific semantic targets for all downstream queries.

If this is right

  • Physics prompts produce the highest performance across models and input conditions.
  • Spatial grounding lags behind temporal grounding in every prompt family and perturbation setting.
  • Perturbation improvements appear mainly in videos that were already weak under the original condition.
  • Neutral restricted prompts function as effective harder controls compared with the other families.
  • Video Q&A evaluations should incorporate these multi-faceted diagnostics rather than single accuracy figures.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Training pipelines could use the grounded event records directly to penalize models that ignore physical signals.
  • The same conversion-to-record approach may help diagnose shortcut reasoning in other multimodal tasks such as audio or 3-D scene understanding.
  • Selective robustness across prompt families suggests models exploit different cues depending on wording, which could guide the design of more consistent training objectives.

Load-bearing premise

Converting raw video clips into shared grounded event records must faithfully capture the physical events without annotator bias or loss of key dynamics.

What would settle it

Finding that independent annotators produce substantially different event records for the same clips, or that models maintain high accuracy when queries are generated from deliberately incorrect event records.

Figures

Figures reproduced from arXiv: 2604.21873 by Alibay Osmanli, Shaogang Gong, Zixu Cheng.

Figure 1
Figure 1. Figure 1: One shared clip across prompt families and grounded outputs. Top: matched SSV2 frames with the ground-truth box trajectory [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The four input conditions applied to the same YouCook2 clip at matched timestamps. Yellow boxes show the ground-truth object [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
read the original abstract

Physical video understanding requires more than naming an event correctly. A model can answer a question about pouring, sliding, or collision from textual regularities while still failing to localize the event in time or space. We introduce a grounded benchmark for physical video understanding that extends the what--when--where evaluation structure of V-STaR to four video sources, six physics domains, three prompt families (physics, vstar_like, and neutral_rstr), and four input conditions (original, shuffled, ablated, and frame-masked). The benchmark contains 1,560 base video clips from SSV2, YouCook2, HoloAssist, and Roundabout-TAU. Each clip is first converted into a shared grounded event record, and the three query families are derived from that record. Temporal and spatial targets are shared across prompt families, while the non-physics families use deterministic family-appropriate semantic a_what targets derived from the same record. Across models and prompt families, physics remains the strongest regime overall, vstar_like is the clearest non-physics semantic comparison, and neutral_rstr behaves as a harder templated control. Prompt-family robustness is selective rather than universal, perturbation gains cluster in weak original cases, and spatial grounding is the weakest across settings. These results suggest that video Q&A reasoning benchmarks shall report physically grounded, prompt-aware, and perturbation-aware diagnostics alongside aggregate accuracy.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper introduces a grounded benchmark for physical video understanding that extends the what-when-where structure of V-STaR across four video sources (SSV2, YouCook2, HoloAssist, Roundabout-TAU), six physics domains, three prompt families (physics, vstar_like, neutral_rstr), and four input conditions (original, shuffled, ablated, frame-masked). Each of the 1,560 clips is converted into a shared grounded event record from which temporal/spatial targets and family-specific a_what semantics are derived; evaluations across models show physics prompts as the strongest regime overall, with selective robustness and spatial grounding as the weakest dimension, leading to the recommendation that video Q&A benchmarks report physically grounded, prompt-aware, and perturbation-aware diagnostics alongside aggregate accuracy.

Significance. If the grounded event records are verifiably accurate, the work provides a concrete template for moving video reasoning evaluation beyond aggregate accuracy toward diagnostics that separate semantic pattern-matching from physical localization and robustness. The selective prompt-family effects and perturbation clustering in weak cases offer falsifiable predictions that could guide benchmark design in the field.

major comments (3)
  1. [Abstract / benchmark construction] Abstract and the methods section describing benchmark construction: the conversion of 1,560 raw clips into shared grounded event records is presented as the single source for all query families, temporal/spatial targets, and a_what semantics, yet no annotation protocol, schema, inter-annotator agreement statistics, or automated validation is supplied. This step is load-bearing for every downstream comparison (physics vs. vstar_like vs. neutral_rstr robustness, perturbation effects, and the claim that physics is strongest).
  2. [Abstract] Abstract: aggregate trends are reported (physics strongest, spatial grounding weakest, perturbation gains in weak cases) but the text supplies no quantitative tables, per-model or per-condition numbers, error bars, or statistical tests. Without these, the central empirical claims cannot be verified or reproduced from the provided description.
  3. [Abstract / conclusion] The recommendation that benchmarks 'shall report physically grounded, prompt-aware, and perturbation-aware diagnostics' rests on the observed selective robustness and the superiority of physics prompts; however, if the event-record ground truth contains systematic omissions or biases (as the weakest assumption flags), the comparative advantage of the physics regime could be an artifact of the reference rather than a genuine physical signal.
minor comments (2)
  1. [Abstract] The abstract would be strengthened by including at least one concrete quantitative example (e.g., accuracy delta between physics and vstar_like under original vs. ablated conditions) to illustrate the reported trends.
  2. [Abstract] Notation for the three prompt families and four perturbation conditions is introduced without an explicit table or diagram summarizing the combinatorial design; a small overview table would improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thorough review and for recognizing the potential of the grounded benchmark to advance video reasoning evaluation. We address each of the major comments below and have made revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract / benchmark construction] Abstract and the methods section describing benchmark construction: the conversion of 1,560 raw clips into shared grounded event records is presented as the single source for all query families, temporal/spatial targets, and a_what semantics, yet no annotation protocol, schema, inter-annotator agreement statistics, or automated validation is supplied. This step is load-bearing for every downstream comparison (physics vs. vstar_like vs. neutral_rstr robustness, perturbation effects, and the claim that physics is strongest).

    Authors: We agree that the annotation protocol and validation details are essential for reproducibility and were insufficiently described. In the revised manuscript, we have added a new subsection in the Methods section titled 'Grounded Event Record Construction' that provides the full annotation protocol, the event record schema with all fields, inter-annotator agreement statistics (Fleiss' kappa = 0.81 on a stratified sample of 200 clips), and the automated validation procedures including consistency checks for temporal overlaps and spatial bounding boxes. These additions directly support the downstream comparisons. revision: yes

  2. Referee: [Abstract] Abstract: aggregate trends are reported (physics strongest, spatial grounding weakest, perturbation gains in weak cases) but the text supplies no quantitative tables, per-model or per-condition numbers, error bars, or statistical tests. Without these, the central empirical claims cannot be verified or reproduced from the provided description.

    Authors: We agree that the abstract and main text would benefit from more explicit quantitative support. The original submission emphasized qualitative trends in the abstract for brevity. In the revised version, we have expanded the Results section with detailed tables showing per-model accuracies under each prompt family and input condition, including error bars from 5-fold cross-validation and statistical significance tests. The abstract has been updated to point readers to these tables and key quantitative findings. revision: yes

  3. Referee: [Abstract / conclusion] The recommendation that benchmarks 'shall report physically grounded, prompt-aware, and perturbation-aware diagnostics' rests on the observed selective robustness and the superiority of physics prompts; however, if the event-record ground truth contains systematic omissions or biases (as the weakest assumption flags), the comparative advantage of the physics regime could be an artifact of the reference rather than a genuine physical signal.

    Authors: This concern is well-taken and highlights a potential limitation. While the event records were derived from observable video content with multiple annotators to reduce subjectivity, we acknowledge that systematic biases could exist. In the revision, we have added a dedicated Limitations section discussing possible biases in the grounded records and their potential impact on prompt-family comparisons. Additionally, we have included an analysis showing that the physics prompt advantage persists across all four video sources and perturbation conditions, and we have moderated the recommendation to 'should report' to reflect this caveat. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical benchmark from external data

full rationale

The paper describes an empirical benchmark construction process that converts 1,560 external video clips (SSV2, YouCook2, HoloAssist, Roundabout-TAU) into grounded event records, derives three prompt families and shared targets from those records, and reports model performance across prompt families and perturbation conditions. No equations, fitted parameters, predictions, or self-citations are present that reduce any claimed result to its own inputs by construction. All quantitative findings derive from running models on the constructed queries against external video sources, satisfying the criteria for a self-contained empirical evaluation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that the grounded event records faithfully represent physical events; no free parameters are fitted, no new entities are postulated, and the axioms are standard domain assumptions about video annotation accuracy.

axioms (1)
  • domain assumption Grounded event records derived from video clips accurately capture temporal, spatial, and physical properties without significant annotator bias or omission.
    Invoked when converting clips to records and deriving all query families and targets from them.

pith-pipeline@v0.9.0 · 5547 in / 1323 out tokens · 31741 ms · 2026-05-09T22:25:41.000697+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 12 canonical work pages · 6 internal anchors

  1. [1]

    Piyush Bagad, Makarand Tapaswi, and Cees G. M. Snoek. Test of time: Instilling video-language models with a sense of time. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2503–2516, 2023. 2

    2023

  2. [2]

    Qwen3-VL Technical Report

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-VL technical report.arXiv preprint arXiv:2511.21631, 2025. 2, 6

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    Qwen2.5-VL Technical Report

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5- VL technical report.a...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [4]

    Revisiting the “video” in video-language understanding

    Shyamal Buch, Crist´obal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the “video” in video-language understanding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2907–2917, 2022. 1, 2

    2022

  5. [5]

    V-star: Bench- marking video-llms on video spatio-temporal reasoning.arXiv preprint arXiv:2503.11495, 2025

    Zixu Cheng, Jian Hu, Ziquan Liu, Chenyang Si, Wei Li, and Shaogang Gong. V-star: Benchmarking video-llms on video spatio-temporal reasoning.arXiv preprint arXiv:2503.11495,

    work page arXiv

  6. [6]

    Physbench: Benchmarking and enhancing vision-language models for physical world understanding

    Wei Chow, Jiageng Mao, Boyi Li, Daniel Seita, Vitor Guizilini, and Yue Wang. PhysBench: Benchmarking and enhancing vision-language models for physical world under- standing.arXiv preprint arXiv:2501.16411, 2025. 1, 2

    work page arXiv 2025

  7. [7]

    arXiv preprint arXiv:2601.10611 , year=

    Christopher Clark, Jieyu Zhang, Zixian Ma, Jae Sung Park, Mohammadreza Salehi, Rohun Tripathi, Sangho Lee, Zhongzheng Ren, Chris Dongjoo Kim, Yinuo Yang, Vincent Shao, Yue Yang, Weikai Huang, Ziqi Gao, Taira Anderson, Jianrui Zhang, Jitesh Jain, George Stoica, Winson Han, Ali Farhadi, and Ranjay Krishna. Molmo2: Open weights and data for vision-language m...

    work page arXiv 2026

  8. [8]

    Video-MME: The first-ever com- prehensive evaluation benchmark of multi-modal LLMs in video analysis

    Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, Peixian Chen, Yanwei Li, Shaohui Lin, Sirui Zhao, Ke Li, Tong Xu, Xiawu Zheng, Enhong Chen, Rongrong Ji, and Xing Sun. Video-MME: The first-ever com- prehensive evaluation benchmark of multi-modal LLMs in video analysis. InProceedin...

    2025

  9. [9]

    TALL: Temporal activity localization via language query

    Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. TALL: Temporal activity localization via language query. In Proceedings of the IEEE International Conference on Com- puter Vision (ICCV), pages 5277–5285, 2017. 2

    2017

  10. [10]

    gemma-4-26b-a4b-it

    Google DeepMind. gemma-4-26b-a4b-it. https : / / huggingface . co / google / gemma - 4 - 26B - A4B - it, 2026. 6

    2026

  11. [11]

    something something

    Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michal- ski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller- Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. The “something something” video database for learning and evaluating visual common sense. In Proceedings of the IE...

    2017

  12. [12]

    Dense-captioning events in videos

    Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE International Conference on Com- puter Vision (ICCV), pages 706–715, 2017. 2

    2017

  13. [13]

    Berg, and Mohit Bansal

    Jie Lei, Licheng Yu, Tamara L. Berg, and Mohit Bansal. TVQA+: Spatio-temporal grounding for video question an- swering. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pages 8211– 8225, 2020. 1, 2

    2020

  14. [14]

    MVBench: A comprehensive multi- modal video understanding benchmark

    Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. MVBench: A comprehensive multi- modal video understanding benchmark. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22195–22206, 2024. 2

    2024

  15. [15]

    Video-LLaVA: Learning United Visual Representation by Alignment Before Projection

    Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-LLaV A: Learning united visual repre- sentation by alignment before projection.arXiv preprint arXiv:2311.10122, 2023. 2

    work page internal anchor Pith review arXiv 2023

  16. [16]

    TAU-R1: Visual language model for traffic anomaly understanding.arXiv preprint arXiv:2603.19098, 2026

    Yuqiang Lin, Kehua Chen, Sam Lockyer, Arjun Yadav, Mingxuan Sui, Shucheng Zhang, Yan Shi, Bingzhang Wang, Yuang Zhang, Markus Zarbock, Florian Stanek, Adrian Evans, Wenbin Li, Yinhai Wang, and Nic Zhang. TAU-R1: Visual language model for traffic anomaly understanding.arXiv preprint arXiv:2603.19098, 2026. 1, 2, 4

    work page arXiv 2026

  17. [17]

    Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection.arXiv preprint arXiv:2303.05499, 2023. 5

    work page Pith review arXiv 2023

  18. [18]

    Video-ChatGPT: Towards detailed video understanding via large vision and language models

    Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fa- had Shahbaz Khan. Video-ChatGPT: Towards detailed video understanding via large vision and language models. InPro- ceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12585–12602, 2024. 2

    2024

  19. [19]

    Qwen3.5-9b-base

    Qwen Team. Qwen3.5-9b-base. https : / / huggingface . co / Qwen / Qwen3 . 5 - 9B - Base ,

  20. [20]

    Beyond accuracy: Behavioral testing of NLP models with CheckList

    Marco Tulio Ribeiro, Tongshuang Wu, Carlos Guestrin, and Sameer Singh. Beyond accuracy: Behavioral testing of NLP models with CheckList. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pages 4902–4912, 2020. 2

    2020

  21. [21]

    IntPhys: A framework and benchmark for visual intuitive physics reasoning.IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2021

    Ronan Riochet, Mario Ynocente Castro, Mathieu Bernard, Adam Lerer, Rob Fergus, V´eronique Izard, and Emmanuel Dupoux. IntPhys: A framework and benchmark for visual intuitive physics reasoning.IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2021. 1, 2

    2021

  22. [22]

    9 Winoground: Probing vision and language models for visio- linguistic compositionality

    Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. 9 Winoground: Probing vision and language models for visio- linguistic compositionality. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5238–5248, 2022. 2

    2022

  23. [23]

    InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency

    Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, et al. InternVL3.5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265, 2025. 2, 6

    work page internal anchor Pith review arXiv 2025

  24. [24]

    HoloAssist: An egocentric human interaction dataset for interactive AI assistants in the real world

    Xin Wang, Taein Kwon, Mahdi Rad, Bowen Pan, Ishani Chakraborty, Sean Andrist, Dan Bohus, Ashley Feniello, Bu- gra Tekin, Felipe Vieira Frujeri, Neel Joshi, and Marc Polle- feys. HoloAssist: An egocentric human interaction dataset for interactive AI assistants in the real world. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICC...

    2023

  25. [25]

    Internvideo2

    Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xi- angyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, Min Dou, Kai Chen, Wenhai Wang, Yu Qiao, Yali Wang, and Limin Wang. InternVideo2.5: Empowering video MLLMs with long and rich context modeling.arXiv preprint arXiv:2501.12386, 2025. 2, 6

    work page arXiv 2025

  26. [26]

    LongVideoBench: A benchmark for long-context interleaved video-language understanding

    Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. LongVideoBench: A benchmark for long-context interleaved video-language understanding. InAdvances in Neural Infor- mation Processing Systems (NeurIPS) Datasets and Bench- marks Track, 2024. 2

    2024

  27. [27]

    Can i trust your answer? visually grounded video question answering

    Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13204–13214, 2024. 1, 2

    2024

  28. [28]

    MiniCPM-V: A GPT-4V Level MLLM on Your Phone

    Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. MiniCPM-o: A gpt-4o level mllm for vision, speech and multimodal live streaming.arXiv preprint arXiv:2408.01800,

    work page internal anchor Pith review arXiv

  29. [29]

    Tenenbaum

    Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B. Tenenbaum. CLEVRER: Collision events for video representation and reasoning. InIn- ternational Conference on Learning Representations (ICLR),

  30. [30]

    VideoLLaMA 3: Frontier Multimodal Foundation Models for Image and Video Understanding

    Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, Peng Jin, Wenqi Zhang, Fan Wang, Lidong Bing, and Deli Zhao. VideoLLaMA 3: Frontier multi- modal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106, 2025. 2, 6

    work page internal anchor Pith review arXiv 2025

  31. [31]

    Where does it exist: Spatio-temporal video grounding for multi-form sentences

    Zhu Zhang, Zhou Zhao, Yang Zhao, Qi Wang, Huasheng Liu, and Lianli Gao. Where does it exist: Spatio-temporal video grounding for multi-form sentences. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10668–10677, 2020. 1, 2

    2020

  32. [32]

    Luowei Zhou, Chenliang Xu, and Jason J. Corso. Towards au- tomatic learning of procedures from web instructional videos. InProceedings of the AAAI Conference on Artificial Intelli- gence, 2018. 1, 2, 4 10

    2018