Research map / Video / Video segmentation and tracking
Understanding Physical Dynamics with Counterfactual World Modeling
Rahul Venkatesh, Honglin Chen, Kevin Feigelis, Daniel M Bear, Khaled Jedoui, Klemen Kotar, Felix J Binder, Wanhee Lee, Sherry Liu, Kevin Smith, Judith E. Fan, Daniel Yamins
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use temporally-factored masking instead of random tube masking (ablation) | CWM · Physion v1.5 OCP | accuracy: 73.2 -> 75.9 | Improved |
| add keypoint patch features to classifier input (ablation) | CWM · Physion v1.5 OCP | accuracy: 73.6 -> 74.4 | Improved |
| add optical flow patches to classifier input (ablation) | CWM · Physion v1.5 OCP | accuracy: 74.4 -> 75.5 | Improved |
| add segment features to classifier input (ablation) | CWM · Physion v1.5 OCP | accuracy: 75.5 -> 75.9 | Improved |
| train 1600 epochs instead of 800 (ablation) | CWM · Physion v1.5 OCP | accuracy: 75.4 -> 75.9 | Improved |
| use 4 context frames instead of 2 during pretraining (ablation) | CWM · Physion v1.5 OCP | accuracy: 75.9 -> 68.5 | Hurt |
| use patch size 8 instead of 16 (ablation) | CWM · Physion v1.5 OCP | accuracy: 74.2 -> 75.9 | Improved |
| raise last-frame mask ratio to 0.99 from 0.90 (ablation) | CWM · Physion v1.5 OCP | accuracy: 75.4 -> 72.5 | Hurt |
| scale CWM from ViT-B to ViT-L | CWM · Physion v1.5 | OCP / OCD accuracy: OCP 75.9 -> 76.1; OCD 89.1 -> 88.7 | Mixed |
| apply temporally-factored CWM instead of VideoMAE for counterfactual motion | DAVIS counterfactual prediction · DAVIS | FID: 166.3 -> 25.4 | Improved |
| overlay ground-truth segment map on GPT-4V prompt frames | GPT-4V · Physion v1.5 | OCP / OCD accuracy: OCP 52.9 -> 58.3; OCD 54.7 -> 67.5 | Improved |
| CWM with 3 frames versus VideoMAE with 16 frames for activity recognition | CWM · Something-Something V2 | accuracy: 54.2 vs 54.7 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VideoMAE* — VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training (NeurIPS 2022) · better on 7 of 7 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- CofactVLA: Deconfounding Vision-Language-Action Models via Counterfactual Intervention (NeurIPS 2026 · similarity 0.85)
- PhysBench: Benchmarking and Enhancing Vision-Language Models for Physical World Understanding (ICLR 2025 · similarity 0.85)
- DynamicVerse: A Physically-Aware Multimodal Framework for 4D World Modeling (NeurIPS 2025 · similarity 0.85)
- RoboScape: Physics-informed Embodied World Model (NeurIPS 2025 · similarity 0.85)
- Learning Physics From Video: Unsupervised Physical Parameter Estimation for Continuous Dynamical Systems (CVPR 2025 · similarity 0.84)
- World Models as Group Actions (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
