Research map / Label-efficient and robust learning / Test-time adaptation
PURA: Parameter Update-Recovery Test-Time Adaption for RGB-T Tracking
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PURA, a gradient-free test-time adaptation framework for RGB-T tracking that updates BatchNorm statistics with test-sample EMA, recovers parameters along principal pseudo-gradient directions via SVD, and adapts momentum.
Problem
Online adaptation of RGB-T trackers to domain shift at test time is needed, but existing TTA methods (entropy minimization, consistency regularization) cause drastic parameter drift in long-term tracking and incur high computational cost.
Approach
Fast BatchNorm statistic updates with test-batch mean/variance (no gradients), plus periodic recovery of parameters along the principal direction of the parameter-change trajectory (pseudo-gradients) obtained by orthogonal/SVD decomposition, with eigenvalue-driven adaptive momentum scaling.
Key results
- On LasHeR→RGBT234, PURA reaches 93.3% MPR and 70.3% MSR versus 90.8%/67.6% for no adaptation and 91.8%/69.3% for Tent.
- On RGBT234 overall, PURA achieves 93.3% MPR and 70.3% MSR, a 5.3% MPR gain over TBSI.
- PURA runs at 42 FPS on GTOT, much faster than Tent (18.5 FPS) while matching its accuracy (95.7% MPR vs 96.0%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add PURA test-time adaptation | ViT-B tracker · RGBT234 | MPR (%): 90.7 -> 93.3 | Helped |
| add fast parameter update (ablation) | ViT-B tracker · LasHeR→RGBT234 | MPR (%): 90.8 -> 91.6 | Helped |
| add fast parameter update (ablation) | ViT-B tracker · LasHeR→RGBT234 | MSR (%): 67.6 -> 69.2 | Helped |
| add parameter decomposition recovery to FPU (ablation) | ViT-B tracker · LasHeR→RGBT234 | MPR (%): 91.6 -> 92.3 | Helped |
| add adaptive momentum scaling to FPU + PDR (ablation) | ViT-B tracker · LasHeR→RGBT234 | MPR (%): 92.3 -> 93.3 | Helped |
| add adaptive momentum scaling to FPU + PDR (ablation) | ViT-B tracker · LasHeR→RGBT234 | MSR (%): 69.6 -> 70.3 | Helped |
| set momentum truncation range to [0.10, 0.11] (ablation) | PURA · LasHeR→RGBT234 | MPR (%): 92.7 -> 93.3; 93.0 -> 93.3; 91.9 -> 93.3 | Helped |
| vary the adaptive momentum value mλ (ablation) | PURA · LasHeR→RGBT234 | MPR (%): 92.9; 92.7; 93.3; 92.2 | Mixed |
| set the adaptive momentum scaling threshold ξ to 1.60 (ablation) | PURA · LasHeR→RGBT234 | MPR (%): 92.9; 92.7; 93.3; 92.2 | Helped |
| vary the recovery learning rates ηµ and ησ (ablation) | PURA · LasHeR→RGBT234 | MPR (%): 93.0; 92.9; 93.2; 92.9; 92.2; 93.2; 93.3 | Mixed |
| set parameter recovery interval K to 3 (ablation) | PURA · LasHeR→RGBT234 | MPR (%): 92.3; 92.8; 92.5; 93.3 | Helped |
| add PURA adaptation | ViT-B tracker · LasHeR testing set, Severity 3 | MPR: 40.1 -> 44.0 | Helped |
Surprising findings
- Tent achieves slightly higher MPR than PURA on GTOT (96.0 vs 95.7), and AdaBN and ETA are faster than PURA.
- On the LasHeR→RGBT210 scenario, PURA matches Tent's MSR (66.8) with only a 0.4-point MPR edge.
Limitations
- Gains over the no-adaptation baseline are modest on some benchmarks, and PURA does not beat Tent on GTOT in MPR.
- Short GTOT sequences limit the benefit of continual updates, as acknowledged by the authors.
- Many hyperparameters (momentum range, threshold ξ, recovery interval K, learning rates) are set empirically.
Experimental setup
Datasets: LasHeR GTOT RGBT210 RGBT234 COCO LaSOT GOT-10k TrackingNet
Benchmarks: GTOT RGBT210 RGBT234 LasHeR
Metrics: MPR MSR PR SR FPS
Baselines: TBSI ViPT MambaVT Tent ETA CoTTA AdaBN MPLT CAT++ HMFT
Builds on: Tent FSTTA AdaBN ETA CoTTA
Closest papers by meaning
- ACAttack: Adaptive Cross Attacking RGB-T Tracker via Multi-Modal Response Decoupling (CVPR 2025 · similarity 0.88)
- Cross-Modal Stealth: A Coarse-to-Fine Attack Framework for RGB-T Tracker (AAAI 2025 · similarity 0.83)
- Temporal Adaptive RGBT Tracking with Modality Prompt (AAAI 2024 · similarity 0.83)
- Cross-modulated Attention Transformer for RGBT Tracking (AAAI 2025 · similarity 0.79)
- CSTrack: Enhancing RGB-X Tracking via Compact Spatiotemporal Features (ICML 2025 · similarity 0.78)
- RGBT Tracking via All-layer Multimodal Interactions with Progressive Fusion Mamba (AAAI 2025 · similarity 0.77)
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.