Research map / Video / Action recognition
How Can Objects Help Action Recognition?
Xingyi Zhou, Anurag Arnab, Chen Sun, Cordelia Schmid
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 |
|---|---|---|---|
| Add object-aware attention while using all tokens | VideoMAE-based ViViT · SomethingElse, SSv2, and Epic-Kitchens | Top-1 accuracy: +2.1, +1.3, and +0.6 percentage points | Improved |
| Use object-guided token sampling instead of uniform random token sampling | Space-time ViViT · SomethingElse, SSv2, and Epic-Kitchens | Top-1 accuracy | Improved |
| Drop background tokens guided by object detections | Space-time ViViT · SomethingElse, SSv2, and Epic-Kitchens | Top-1 accuracy: 0.2 points | Improved |
| Use object-guided sampling with ground-truth detections | Space-time ViViT · SomethingElse and SSv2 | Top-1 accuracy: 2.1% gain on both Something datasets | Improved |
| Retain object-guided tokens and add two-view testing at 50% tokens | ObjectViViT · SomethingElse, SSv2, and Epic-Kitchens | Top-1 accuracy | Improved |
| Retain 30%, 40%, and 60% of tokens, respectively | ObjectViViT · SomethingElse, SSv2, and Epic-Kitchens | Top-1 accuracy: 30%, 40%, and 60% of tokens | No clear effect |
| Retain only foreground/object tokens and remove all background tokens (ablation) | Object-guided token sampling · SSv2 | Top-1 accuracy | Hurt |
| Apply token sampling before transformer layers rather than at layer 1 or layer 6 (ablation) | Object-guided token sampling · SSv2 | Top-1 accuracy: 66.1 vs 66.0 and 65.9 | No clear effect |
| Use early token sampling rather than sampling at layer 1 or 6 (ablation) | Object-guided token sampling · SSv2 | gFLOPs: 80 vs 88 and 130 | Improved |
| Replace RoIAlign or binary block-mask pooling with center-heatmap-weighted pooling (ablation) | Object-aware attention · SSv2 | Top-1 accuracy: 67.2 -> 67.4 | Improved |
| Add tracking identity information rather than using no tracking (ablation) | Object-aware attention · SSv2 | Top-1 accuracy: 67.0 -> 67.4 | Improved |
| Use identity attention rather than a shared identity embedding (ablation) | Object-aware attention · SSv2 | Top-1 accuracy: 67.4 vs 67.4 | 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 2 of 2 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
- Attend and Interact: Higher-Order Object Interactions for Video Understanding (CVPR 2018 · similarity 0.89)
- Scenes-Objects-Actions: A Multi-Task, Multi-Label Video Dataset (ECCV 2018 · similarity 0.88)
- Bringing Image Scene Structure to Video via Frame-Clip Consistency of Object Tokens (NeurIPS 2022 · similarity 0.88)
- Object Priors for Classifying and Localizing Unseen Actions (IJCV 2021 · similarity 0.88)
- VideoOrion: Tokenizing Object Dynamics in Videos (ICCV 2025 · similarity 0.87)
- Action Detail Matters: Refining Video Recognition with Local Action Queries (CVPR 2025 · similarity 0.86)
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.
