Research map / Video / Action recognition
Joint Discovery of Object States and Manipulation Actions
Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, Simon Lacoste-Julien
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 the at-least-one-state constraint instead of exactly-one (ablation) | state-only discovery · State discovery (weakly supervised clips, average over 7 actions) | precision: 0.38 -> 0.44 | Improved |
| add ordering and non-overlap constraints with learned features instead of random features (ablation) | state discovery · State discovery (weakly supervised clips, average over 7 actions) | precision: 0.35 -> 0.44 | Improved |
| add joint modeling with actions to state-only discovery | state discovery · State discovery (weakly supervised clips, average over 7 actions) | precision: 0.44 -> 0.48 | Improved |
| add object detection score to the joint objective | joint model · State discovery (weakly supervised clips, average over 7 actions) | precision: 0.48 -> 0.51 | Improved |
| use ground-truth action features instead of predicted actions (ablation) | joint model · State discovery (weakly supervised clips, average over 7 actions) | precision: 0.48 -> 0.51 | No clear effect |
| add object-state term to action localization (joint model) over salient action only | action localization · Action localization (weakly supervised clips, average over 7 actions) | precision: 0.20 -> 0.55 | Improved |
| constrain action to lie within first and last frame of object detection (ablation) | action localization baseline · Action localization (weakly supervised clips, average over 7 actions) | precision: 0.20 -> 0.32 | Improved |
| use ground-truth state labels as tracklet features (ablation) | joint action localization · Action localization (weakly supervised clips, average over 7 actions) | precision: 0.55 -> 0.59 | Improved |
| add joint modeling over state-only with detection scores in the wild | state discovery · State discovery (weakly supervised clips, average over 7 actions) in the wild | precision: 0.32 -> 0.36 | Improved |
| use curated clips instead of automatically retrieved clips (ablation) | joint model · State discovery in the wild | precision: 0.36 -> 0.62 | Improved |
| use curated clips instead of automatically retrieved clips (ablation) | joint model · Action localization in the wild | precision: 0.24 -> 0.44 | Improved |
| add joint modeling to action-only localization in the wild | action localization · Action localization in the wild | precision: 0.11 -> 0.24 | Improved |
Closest papers by meaning
- Look for the Change: Learning Object States and State-Modifying Actions From Untrimmed Web Videos (CVPR 2022 · similarity 0.90)
- Multi-Task Learning of Object States and State-Modifying Actions From Web Videos (TPAMI 2024 · similarity 0.89)
- SAGE: A Unified Framework for Generalizable Object State Recognition with State-Action Graph Embedding (NeurIPS 2025 · similarity 0.89)
- InfoCon: Concept Discovery with Generative and Discriminative Informativeness (ICLR 2024 · similarity 0.88)
- MOSCATO: Predicting Multiple Object State Change Through Actions (ICCV 2025 · similarity 0.88)
- Learning Object State Changes in Videos: An Open-World Perspective (CVPR 2024 · similarity 0.87)
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.
