Research map / 3D vision / 3D generation
MITracker: Multi-View Integration for Visual Object 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.
Introduces the MVTrack multi-view object tracking dataset (234K frames, 27 objects) and MITracker, which lifts ViT view features into a BEV-guided 3D feature volume and uses spatial-enhanced attention to refine per-view tracking.
Problem
Multi-view object tracking for class-agnostic objects is limited by the lack of comprehensive multi-view datasets with training splits and by weak cross-view integration, leaving occlusion and target loss unresolved.
Approach
Project per-view 2D features into a 3D feature volume compressed to a bird's eye view plane with BEV supervision, then feed the fused 3D token back via spatial-enhanced attention to refine each view's tracking output.
Key results
- MITracker reaches 68.57% AUC on MVTrack single-view setting, about 5% above ODTrack (63.36%).
- MITracker reports 79.2% recovery rate within 10 frames after target disappears versus 56.7% for SAM2Long and ODTrack.
- In the multi-view setting on MVTrack, MITracker surpasses second-best OSTrack by about 26% in PNorm (88.77 vs 77.72).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MITracker instead of the strongest single-view baseline | visual object tracking · MVTrack single-view | AUC: 63.36 -> 68.57 | Helped |
| use MITracker multi-view integration instead of post-fusion baseline tracking | multi-view tracking · MVTrack multi-view | PNorm: 65.19 -> 91.87 | Helped |
| use MITracker instead of the strongest single-view baseline | visual object tracking · GMTD single-view | AUC: 63.89 -> 65.96 | Helped |
| add BEV loss (ablation) | MITracker · MVTrack multi-view | AUC: 63.99 -> 69.64 | Helped |
| add spatial attention (ablation) | MITracker with BEV loss · MVTrack multi-view | AUC: 69.64 -> 71.13 | Helped |
| add BEV loss (ablation) | MITracker · MVTrack multi-view | PNorm: 82.82 -> 89.85 | Helped |
| add spatial attention (ablation) | MITracker with BEV loss · MVTrack multi-view | PNorm: 89.85 -> 91.87 | Helped |
| add BEV loss (ablation) | MITracker · MVTrack multi-view | P: 75.00 -> 82.01 | Helped |
| add spatial attention (ablation) | MITracker with BEV loss · MVTrack multi-view | P: 82.01 -> 83.95 | Helped |
| use MITracker for recovery after target disappearance | visual object tracking · MVTrack | recovery rate: 56.7% to 79.2% | Helped |
Surprising findings
- Applying post-fusion of single-view predictions via geometric projection degrades performance of all single-view trackers, indicating a distribution gap across models.
Limitations
- MVTrack contains only indoor scenes, limiting generalization to outdoor settings.
- MITracker depends on precise camera calibration for multi-view fusion.
- Baselines are single-view methods evaluated with post-fusion, so the comparison is not like-for-like.
Experimental setup
Datasets: MVTrack GOT-10K GMTD
Benchmarks: MVTrack GMTD
Metrics: AUC PNorm Precision recovery rate max continuous tracking length restart count
Baselines: DiMP PrDiMP TrDiMP MixFormer OSTrack GRM SeqTrack ARTrack HIPTrack EVPTrack AQATrack ODTrack
Builds on: ODTrack CenterNet DINOv2 LaSOT annotation protocol BEV feature projection
Closest papers by meaning
- Cross-View Referring Multi-Object Tracking (AAAI 2025 · similarity 0.86)
- Exploring Enhanced Contextual Information for Video-Level Object Tracking (AAAI 2025 · similarity 0.85)
- CoMatcher: Multi-View Collaborative Feature Matching (CVPR 2025 · similarity 0.84)
- HFF-Tracker: A Hierarchical Fine-grained Fusion Tracker for Referring Multi-Object Tracking (AAAI 2025 · similarity 0.84)
- VastTrack: Vast Category Visual Object Tracking (NeurIPS 2024 · similarity 0.84)
- ADA-Track: End-to-End Multi-Camera 3D Multi-Object Tracking with Alternating Detection and Association (CVPR 2024 · 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.