Research map / Detection and segmentation / Tracking
ODTrack: Online Dense Temporal Token Learning for Visual 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 ODTrack, a video-level visual tracker that densely propagates a temporal token sequence across frames in an online auto-regressive manner, avoiding explicit online update strategies.
Problem
Visual trackers typically match sparse reference-search image pairs offline, so they capture limited temporal correlations and motion/trajectory information across frames.
Approach
Extends the transformer tracker input from image pairs to arbitrary-length video clips and adds a temporal token per frame that compresses target localization information and is propagated to the next frame as a prompt; two attention variants (concatenated and separated token attention) implement this compression-propagation.
Key results
- ODTrack-L reaches 74.0% AUC on LaSOT and 78.2% AO on GOT-10k (trained on GOT-10k only), setting new SOTA over ARTrack (72.6% and 75.5%).
- ODTrack-B achieves 0.581 EAO on VOT2020 and runs at 32 fps on a 2080Ti with 92M params and 73G FLOPs, versus SeqTrack's 11 fps.
- Removing token propagation drops LaSOT AUC by 1.2% (71.0 vs 72.2/72.8 with separate/concatenated attention).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ODTrack-B instead of ARTrack | ODTrack-B · GOT10K | AO: 75.5 -> 77.0 | Helped |
| use ODTrack-B instead of ARTrack | ODTrack-B · LaSOT | AUC: 71.5 -> 73.2 | Helped |
| use ODTrack-B instead of SeqTrack | ODTrack-B · TrackingNet | P_Norm: 88.8 -> 90.1 | Helped |
| use ODTrack-B instead of ARTrack | ODTrack-B · LaSOT_ext | AUC: 51.9 -> 52.4 | Helped |
| use ODTrack-B instead of Mixformer | ODTrack-B · VOT2020 | EAO: 0.535 -> 0.581 | Helped |
| use ODTrack-B instead of ARTrack | ODTrack-B · TNL2K | AUC: 59.8 -> 60.9 | Helped |
| use ODTrack-B instead of TransInMo | ODTrack-B · OTB100 | AUC: 71.1 -> 72.3 | Helped |
| add temporal token propagation to video-level sampling (ablation) | ODTrack · LaSOT | AUC: 71.0 -> 72.2 | Helped |
| use concatenated instead of separated token attention (ablation) | ODTrack · LaSOT | AUC: 72.2 -> 72.8 | Helped |
| increase video sequence length from 2 to 3 (ablation) | ODTrack · LaSOT | AUC: 72.8 -> 73.1 | Helped |
| increase video sequence length from 3 to 4 (ablation) | ODTrack · LaSOT | AUC: 73.1 -> 72.5 | Hurt |
| increase video sequence length from 4 to 5 (ablation) | ODTrack · LaSOT | AUC: 72.5 -> 72.0 | Hurt |
Surprising findings
- Longer search clips do not monotonically help: sequence length 3 is best, and lengths 4 and 5 degrade performance.
Limitations
- Video-level modeling is a global approximation constrained by GPU resources.
- Not cost-effective; computational complexity remains high for long video modeling.
- VOT2020 mask evaluation depends on the external Alpha-Refine post-processing network.
Experimental setup
Datasets: LaSOT GOT-10k TrackingNet COCO
Benchmarks: GOT-10k LaSOT TrackingNet LaSOT_ext VOT2020 TNL2K OTB100
Metrics: AO SR0.5 SR0.75 AUC P_Norm Precision EAO Accuracy Robustness
Baselines: OSTrack SeqTrack ARTrack Mixformer STARK SBT VideoTrack TransInMo AiATrack GRM
Builds on: OSTrack STARK MAE ViT
Closest papers by meaning
- Less Is More: Token Context-Aware Learning for Object Tracking (AAAI 2025 · similarity 0.91)
- SPMTrack: Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking (CVPR 2025 · similarity 0.88)
- UTPTrack: Towards Simple and Unified Token Pruning for Visual Tracking (CVPR 2026 · similarity 0.88)
- Track-On: Transformer-based Online Point Tracking with Memory (ICLR 2025 · similarity 0.87)
- Exploring Enhanced Contextual Information for Video-Level Object Tracking (AAAI 2025 · similarity 0.87)
- Robust Tracking via Mamba-based Context-aware Token Learning (AAAI 2025 · similarity 0.87)
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.