Research map / Detection and segmentation / Tracking
TAPTRv3: Spatial and Temporal Context Foster Robust Tracking of Any Point in Long Video
Jinyuan Qu, Hongyang Li, Shilong Liu, Tianhe Ren, Zhaoyang Zeng, Lei Zhang
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 |
|---|---|---|---|
| replace RNN-like long-temporal modeling with long-temporal attention (ablation) | TAPTRv2 · TAP-Vid-Kinetics (ablation) | AJ: 44.5 -> 47.8 | Improved |
| add visibility-aware reweighting to long-temporal attention (ablation) | TAPTRv3 · TAP-Vid-Kinetics (ablation) | AJ: 47.8 -> 48.8 | Improved |
| remove sliding window (ablation) | TAPTRv3 · TAP-Vid-Kinetics (ablation) | AJ: 48.8 -> 49.5 | Improved |
| supervise only visible points' positions (ablation) | TAPTRv3 · TAP-Vid-Kinetics (ablation) | AJ: 49.5 -> 51.1 | Improved |
| add context-aware cross-attention (CCA) (ablation) | TAPTRv3 · TAP-Vid-Kinetics (ablation) | AJ: 51.1 -> 52.9 | Improved |
| use every-two-point patch similarity instead of element-wise (ablation) | CCA · TAP-Vid-Kinetics (ablation) | AJ: 51.3 -> 52.9 | Improved |
| update spatial context features with VLTA or MLP during tracking (ablation) | TAPTRv3 · TAP-Vid-Kinetics (ablation) | AJ: 52.9 -> 51.2 (VLTA), 51.7 (MLP) | Hurt |
| use global matching at every frame for decoder initial position (ablation) | TAPTRv3 · TAP-Vid-Kinetics (ablation) | AJ: 52.9 -> 51.1 | Hurt |
| trigger global matching only on scene cuts (ablation) | TAPTRv3 · TAP-Vid-Kinetics (full) | AJ: 54.9 -> 55.1 | Improved |
| trigger global matching on scene cuts (ablation) | TAPTRv3 · Kinetics scene-cut subset | AJ: 55.3 -> 55.8 | Improved |
| use N=3 (9) context features instead of N=5 (25) or N=1 (ablation) | CCA · TAP-Vid-Kinetics (ablation) | AJ: N=1 51.3, N=3 52.9, N=5 52.2 | Mixed |
| increase VLTA memory size up to all past frames (ablation) | TAPTRv3 · TAP-Vid-Kinetics | AJ: 51.9 (12) -> 54.9 (all past) | Improved |
Closest papers by meaning
- TAPTRv2: Attention-based Position Update Improves Tracking Any Point (NeurIPS 2024 · similarity 0.90)
- TAPIP3D: Tracking Any Point in Persistent 3D Geometry (NeurIPS 2025 · similarity 0.89)
- TAPVid-3D: A Benchmark for Tracking Any Point in 3D (NeurIPS 2024 · similarity 0.88)
- A Distractor-Aware Memory for Visual Object Tracking with SAM2 (CVPR 2025 · similarity 0.86)
- Re-thinking Temporal Search for Long-Form Video Understanding (CVPR 2025 · similarity 0.85)
- SALOVA: Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis (CVPR 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.
