Research map / Video / Action recognition
Trajectory-aligned Space-time Tokens for Few-shot Action Recognition
Pulkit Kumar, Namitha Padmanabhan, Luke Luo, Sai Saketh Rambhatla, Abhinav Shrivastava
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 point trajectories (uniform temporal re-initialization) to DINO tokens (ablation) | TATs · Kinetics 5-way 1-shot | accuracy (%): 80.2 -> 81.9 | Improved |
| initialize points only in first frame instead of no points (ablation) | TATs · Kinetics 5-way 1-shot | accuracy (%): 80.2 -> 78.6 | Hurt |
| reinitialize points uniformly over time instead of first frame only (ablation) | TATs · SSv2-Full 5-way 1-shot | accuracy (%): 56.0 -> 57.7 | Improved |
| increase sampled points from 32 to 256 (ablation) | TATs · SSv2-Full 5-way 1-shot | accuracy (%): 53.4 -> 57.7 | Improved |
| halve points from 256 to 128 (ablation) | TATs · Kinetics 5-way 1-shot | accuracy (%): 81.9 -> 81.6 | No clear effect |
| increase grid size from 16x16 to 25x25 (ablation) | TATs · SSv2-Full 5-way 1-shot | accuracy (%): 57.7 -> 58.1 | No clear effect |
| reduce grid size to 9x9 (ablation) | TATs · SSv2-Full 5-way 1-shot | accuracy (%): 57.7 -> 56.0 | Hurt |
| swap CoTrackerV2 for TAPIR or PIPS++ (ablation) | TATs · SSv2-Full 5-way 1-shot | accuracy (%): 57.7 -> 56.9 (TAPIR), 57.3 (PIPS++) | No clear effect |
| use trajectory length or HOD sampling instead of random sampling (ablation) | TATs · SSv2-Full 5-way 1-shot | accuracy (%): 57.7 -> 57.4 (length), 57.1 (HOD) | No clear effect |
| add point trajectories on top of DINO tokens (ablation) | TATs · FineGym 1-shot | accuracy (%): 77.6 -> 81.8 | Improved |
| use TATs instead of MoLo on HMDB-51 | TATs · HMDB-51 5-way 1-shot | accuracy (%): 65.1 -> 60.0 (vs CCLN) | Hurt |
| use only 2 input frames | TATs · SSv2-Full 5-way 1-shot | accuracy (%) | Hurt |
Closest papers by meaning
- Trokens: Semantic-Aware Relational Trajectory Tokens for Few-Shot Action Recognition (ICCV 2025 · similarity 0.91)
- TA2N: Two-Stage Action Alignment Network for Few-Shot Action Recognition (AAAI 2022 · similarity 0.88)
- Temporal Alignment-Free Video Matching for Few-shot Action Recognition (CVPR 2025 · similarity 0.87)
- TrAction: Action Recognition with Sparse Trajectories (NeurIPS 2026 · similarity 0.87)
- Learning Causal Domain-Invariant Temporal Dynamics for Few-Shot Action Recognition (ICML 2024 · similarity 0.86)
- Frame Order Matters: A Temporal Sequence-Aware Model for Few-Shot Action Recognition (AAAI 2025 · similarity 0.85)
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.
