Research map / Label-efficient and robust learning / Test-time adaptation
Stabilizing Real-World Visual Active Tracking with Action-Smooth Test-Time Adaptation
Haowei Sun, Shiteng Zhang, Jinwu Hu, Kaining Chen, Mingkui Tan
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 Critic Value Maximization (ablation) | VATA (EVT base) · UnrealCV average | SR: 0.80 -> 0.85 | Improved |
| add Action Smoothness Regularization (ablation) | VATA (EVT base) · UnrealCV average | SR: 0.82 -> 0.85 | Improved |
| apply VATA test-time adaptation to base tracker | EVT · UnrealCV average | SR: 0.79 -> 0.85 | Improved |
| use VATA instead of entropy-based TARL | VATA vs TARL · UnrealCV average | SR: 0.82 -> 0.85 | Improved |
| apply VATA (critic value maximization only, discrete actions) | GC-VAT · DAT average | TSR: 0.62 -> 0.67 | Improved |
| use critic value maximization instead of entropy minimization (TARL) | VATA vs TARL · DAT average | TSR: 0.64 -> 0.67 | Improved |
| apply VATA to base tracker | EVT · EVT-Bench average | SR: 15.1% -> 22.4% | Improved |
| apply VATA versus runner-up TTA on distractor task | VATA vs TARL · EVT-Bench DT | SR: 6.8% -> 11.2% | Improved |
| add action smoothness regularization to imitation-learning tracker | Open-TrackVLA · EVT-Bench DT | SR: 11.6% -> 16.9% | Improved |
| deploy VATA on a real drone | EVT · Real drone (10 episodes) | TSR: 0.2 -> 0.8 | Improved |
| vary critic threshold Q0 in {5,10,20} and jump threshold delta in {0.1,0.2,0.3} (ablation) | VATA · UnrealCV average | AR: AR variation less than 3.3% | No clear effect |
| filter samples by critic value rather than output entropy (ablation) | VATA · DAT (GC-VAT) | success vs loss gap: critic 9.12 -> 3.10 on target loss; entropy 0.580 vs 0.560 | Improved |
Closest papers by meaning
- Long-term Embodied Visual Tracking with Lightweight Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- Instance-level Visual Active Tracking with Occlusion-Aware Planning (CVPR 2026 · similarity 0.88)
- Learning Adaptive and View-Invariant Vision Transformer for Real-Time UAV Tracking (ICML 2024 · similarity 0.86)
- Open-World Drone Active Tracking with Goal-Centered Rewards (NeurIPS 2025 · similarity 0.85)
- DiT4DiT: Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control (NeurIPS 2026 · similarity 0.85)
- SyncWorld: Visual Calibration Enables World Models as Zero-Shot Simulators (NeurIPS 2026 · similarity 0.84)
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.
