Research map / Video / Action recognition
Universal Prototype Transport for Zero-Shot Action Recognition and Localization
Pascal Mettes
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 universal prototype transport with interpolation ratio 0.5 | universal action model · UCF-101 (101 actions) | top-1 accuracy: 39.2 -> 42.4 | Improved |
| interpolate halfway along geodesic instead of using target prototypes directly (ablation) | universal prototype transport · UCF-101 (101 actions) | top-1 accuracy: 40.1 -> 42.4 | Improved |
| use enough clusters (1,000) for the test-video distribution (ablation) | universal prototype transport · UCF-101 | top-1 accuracy | Improved |
| add universal prototype transport | action model 1 · UCF-101 | top-5 accuracy: 60.8 -> 69.2 | Improved |
| add universal prototype transport | action model 2 · UCF-101 | top-1 accuracy: 46.3 -> 47.0 | Improved |
| add object-based transport with weights and filtering to 1,000 objects | universal object model · UCF-101 | top-1 accuracy: 29.9 -> 31.6 | Improved |
| use vanilla uniform optimal transport over all objects (ablation) | universal object model · UCF-101 | top-1 accuracy: 29.9 -> 30.1 | No clear effect |
| keep only top 500 objects (too strict threshold) (ablation) | object transport · UCF-101 | top-1 accuracy: 31.6 -> 30.2 | Hurt |
| add inverse action weighting and transductive object weighting (ablation) | object transport · UCF-101 | top-1 accuracy: 29.8 -> 31.6 | Improved |
| fuse universal action and object models equally with transport | action transport · UCF-101 | top-1 accuracy: 42.4 -> 47.9 | Improved |
| add transport to action model and fuse with objects | action model · HMDB51 (51 actions) | top-1 accuracy: 24.9 -> 28.1 | Improved |
| apply transport under test-time class imbalance (ratio 0.01) (ablation) | universal prototype transport · UCF-101 long-tailed | per-class accuracy: 42.0 -> 41.6 | No clear effect |
Closest papers by meaning
- Spatial-Aware Object Embeddings for Zero-Shot Localization and Classification of Actions (ICCV 2017 · similarity 0.90)
- Object Priors for Classifying and Localizing Unseen Actions (IJCV 2021 · similarity 0.89)
- Transductive Zero-Shot Action Recognition by Word-Vector Embedding (IJCV 2017 · similarity 0.88)
- AVA: A Video Dataset of Spatio-Temporally Localized Atomic Visual Actions (CVPR 2018 · similarity 0.88)
- Test-Time Zero-Shot Temporal Action Localization (CVPR 2024 · similarity 0.87)
- Building a Multi-modal Spatiotemporal Expert for Zero-shot Action Recognition with CLIP (AAAI 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.
