Research map / Label-efficient and robust learning / Domain adaptation and generalization
LVPTrack: High Performance Domain Adaptive UAV Tracking with Label Aligned Visual Prompt Tuning
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 LVPTrack, a domain-adaptive UAV tracker that uses a mean-teacher ViT encoder trained on synthesized foggy and nighttime video, plus target-aware pseudo-label voting and a dynamic aggregated prompt module, reaching 110 FPS with state-of-the-art accuracy on haze and dark aerial benchmarks.
Problem
UAV object tracking degrades severely under foggy and nighttime conditions due to domain gap, and existing domain-adaptive UAV trackers produce noisy pseudo labels that misalign target localization.
Approach
Combine atmospheric-scattering haze synthesis and CoGAN night translation with a mean-teacher ViT tracker, refine teacher pseudo labels via score-map-based soft voting over top-k peaks, and adapt via a learned dynamic aggregated prompt on frozen backbone features.
Key results
- On UAV123-Haze, LVPTrack reaches 61.01 AUC and 81.12 precision, 3.92 AUC above the LiteTrack-B8 baseline (57.09).
- On DTB70-Dark, LVPTrack achieves 64.29 AUC and 82.36 precision, the best among compared trackers.
- Ablation on UAV123-Haze shows PLV adds +3.32 AUC over Baseline-Haze and the prompt module adds further gains to 61.01 AUC; the tracker runs at 110 FPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the synthetic haze domain dataset to source-domain training (ablation) | LVPTrack training · UAV123-Haze | AUC: 57.09 -> 59.03 | Helped |
| Add pseudo-label voting (ablation) | Baseline-Haze · UAV123-Haze | AUC: 59.03 -> 60.41 | Helped |
| Add the prompt module (ablation) | PLV · UAV123-Haze | AUC: 59.03 -> 61.01 | Helped |
| Average the top K boxes by classification score instead of using PLV (ablation) | LVPTrack pseudo-label selection · UAV123-Haze | AUC: 59.03 -> 58.25 | Hurt |
| Filter top K boxes to those nearest the highest-scoring point (ablation) | LVPTrack pseudo-label selection · UAV123-Haze | AUC: 59.03 -> 61.01 | Helped |
| Update EMA each epoch rather than every 5 epochs (ablation) | LVPTrack · UAV123-Haze | AUC: 60.21 -> 61.01 | Helped |
| Update EMA each epoch rather than each batch (ablation) | LVPTrack · UAV123-Haze | AUC: 59.44 -> 61.01 | Helped |
| Set the synthetic dataset sampling ratio to 1:1:1:1:2 (ablation) | LVPTrack training · UAV123-Haze | AUC: 60.37 -> 61.01 | Helped |
| Set the synthetic dataset sampling ratio to 1:1:1:1:2 rather than 1:1:1:1:1 (ablation) | LVPTrack training · UAV123-Haze | AUC: 60.40 -> 61.01 | Helped |
| Set the synthetic dataset sampling ratio to 1:1:1:1:2 rather than 1:1:1:1:4 (ablation) | LVPTrack training · UAV123-Haze | AUC: 60.33 -> 61.01 | Helped |
| Use LVPTrack rather than SimTrack-B/16 | UAV aerial tracking · UAV123-Haze | AUC: 60.98 -> 61.01 | Helped |
| Use LVPTrack rather than SimTrack-B/16 | UAV aerial tracking · DTB70-Haze | AUC: 57.03 -> 63.51 | Helped |
Surprising findings
- Top-K averaging of pseudo-label boxes raised precision but reduced AUC (-0.78), showing merging boxes from similar targets hurts.
Limitations
- Synthetic haze and night data may not fully match real adverse-weather appearance; evaluation on real foggy data is absent.
- Pseudo-label voting relies on the teacher's score map quality and hyperparameters such as top-k and sigma, which are tuned empirically.
Experimental setup
Datasets: LaSOT TrackingNet COCO GOT-10k GOT-10k-Extreme UAV123 DTB70 NAT2021-test UAVDark70
Benchmarks: UAV123-Haze DTB70-Haze UAV123-Dark DTB70-Dark NAT2021-test UAVDark70
Metrics: AUC precision speed (FPS)
Baselines: OSTrack-256 LiteTrack-B8 SimTrack-B/16 SeqTrack-B256 TCTrack HiFT UDAT-BAN UDAT-CAR SiamTPN DiMP
Builds on: OSTrack LiteTrack VPT Mean teacher Cross-domain adaptive teacher
Closest papers by meaning
- Learning Adaptive and View-Invariant Vision Transformer for Real-Time UAV Tracking (ICML 2024 · similarity 0.87)
- Learning Occlusion-Robust Vision Transformers for Real-Time UAV Tracking (CVPR 2025 · similarity 0.85)
- Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking (CVPR 2025 · similarity 0.85)
- LRR: Language-Driven Resamplable Continuous Representation against Adversarial Tracking Attacks (ICLR 2024 · similarity 0.84)
- Dynamic Semantic-Aware Correlation Modeling for UAV Tracking (NeurIPS 2025 · similarity 0.84)
- Open-World Drone Active Tracking with Goal-Centered Rewards (NeurIPS 2025 · 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.