Research map / Label-efficient and robust learning / Recognition and long-tail
RTracker: Recoverable Tracking via PN Tree Structured Memory
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 RTracker, which pairs a tracker and a detector through a Positive-Negative tree-structured memory and walking rules to detect target loss and recover tracking.
Problem
Visual object tracking fails to recover after target loss caused by full occlusion, out-of-view, or tracking failure; fixed-threshold presence checks generalize poorly.
Approach
A PN tree memory stores chronological positive and negative target samples, and walking rules classify a test node by relative cosine similarity to these support samples, instead of fixed thresholds; control flows then switch between tracker and global detector.
Key results
- On VideoCube, RTracker-L reaches 69.6 SUC/AUC versus 67.2 for the base tracker MixViT-L, with NP 81.5 vs 78.7.
- On LaSOT, RTracker-L achieves 74.7 AUC and 84.5 NP, improving over MixViT-L by about 1.4 AUC.
- In recovery evaluation on LaSOT, RTracker recovers 80% of lost targets versus 63% for MixViT, 54% for OSTrack and 53% for SeqTrack.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| associate the tracker and detector using PN-tree target-state prediction (ablation) | MixViT-L tracker and MITS detector · VideoCube | NP: 78.7 -> 81.5 | Helped |
| associate the tracker and detector using PN-tree target-state prediction (ablation) | MixViT-L tracker and MITS detector · LaSOText | NP: 61.6 -> 65.5 | Helped |
| associate the tracker and detector using PN-tree target-state prediction (ablation) | MixViT-L tracker and MITS detector · TNL2K | NP: 75.5 -> 76.9 | Helped |
| replace tracker-only tracking with detector-only tracking (ablation) | RTracker · VideoCube | AUC: 69.6 -> 46.6 | Hurt |
| replace tracker-only tracking with detector-only tracking (ablation) | RTracker · LaSOText | AUC: 54.9 -> 50.3 | Hurt |
| replace tracker-only tracking with detector-only tracking (ablation) | RTracker · TNL2K | AUC: 60.6 -> 55.5 | Hurt |
| replace PN-tree state prediction with a fixed confidence threshold (ablation) | RTracker · VideoCube | AUC: 60.9 -> 69.6 | Hurt |
| replace PN-tree state prediction with a fixed confidence threshold (ablation) | RTracker · LaSOText | AUC: 50.7 -> 54.9 | Hurt |
| replace PN-tree state prediction with a fixed confidence threshold (ablation) | RTracker · TNL2K | AUC: 56.5 -> 60.6 | Hurt |
| remove PN-tree walking rules (ablation) | RTracker · VideoCube | AUC: 62.6 -> 69.6 | Hurt |
| remove PN-tree walking rules (ablation) | RTracker · LaSOText | AUC: 49.0 -> 54.9 | Hurt |
| remove PN-tree walking rules (ablation) | RTracker · TNL2K | AUC: 57.4 -> 60.6 | Hurt |
Surprising findings
- Ablation shows the fixed-threshold variant of the state predictor underperforms the base tracker on VideoCube (AUC 60.9 vs 67.2), indicating thresholds hurt.
- RTracker runs at 0.75 times the speed of its baseline tracker due to the extra state evaluation and detector integration.
Limitations
- Runs at 0.75x the baseline speed due to extra computation.
- Fails when the target reappears with a completely different appearance amid similar distractors (failure case).
- Relies on a separate detector (MITS) and a tracker, increasing system complexity.
Experimental setup
Datasets: LaSOT NIGHTS GOT-10k
Benchmarks: VideoCube LaSOT LaSOText TNL2K GOT-10k
Metrics: normalized precision AUC success rate precision AO SR0.5 SR0.75 recovery success rate
Baselines: MixViT-L SeqTrack-L OSTrack ARTrack-L UNINEXT-H MITS CiteTracker DropTrack TLD
Builds on: MixViT MITS DreamSim
Closest papers by meaning
- When Trackers Fail: VLM-Guided Verification and Recovery for Robust Video Object Segmentation (NeurIPS 2026 · similarity 0.85)
- Track-On: Transformer-based Online Point Tracking with Memory (ICLR 2025 · similarity 0.85)
- MIMTrack: In-Context Tracking via Masked Image Modeling (AAAI 2025 · similarity 0.84)
- TGFormer: Transformer with Track Query Group for Multi-Object Tracking (AAAI 2025 · similarity 0.84)
- A Distractor-Aware Memory for Visual Object Tracking with SAM2 (CVPR 2025 · similarity 0.84)
- UTPTrack: Towards Simple and Unified Token Pruning for Visual Tracking (CVPR 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.