Research map / Detection and segmentation / Tracking
Cross-View Referring Multi-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 cross-view referring multi-object tracking (CRMOT), builds the CRTrack benchmark from DIVOTrack and CAMPUS with language descriptions, and proposes CRTracker, an end-to-end baseline combining CrossMOT and APTM.
Problem
Referring multi-object tracking in single views often fails because object appearances are invisible or occluded, so language-matched objects are misidentified; the paper targets tracking language-matched objects while keeping identity consistent across overlapping camera views.
Approach
Extends RMOT to multiple overlapping views so that object appearance can be observed from several cameras, and couples a CenterNet-based cross-view MOT network with APTM attribute-prompt text matching plus a prediction module that filters trajectories by fused per-view scores.
Key results
- On the CRTrack in-domain test set, CRTracker reaches 54.88% CVRIDF1 and 35.97% CVRMA across all scenes, versus 23.43% and 10.14% for the best TempRMOT variant.
- On the cross-domain test set (CAMPUS scenes), CRTracker reports 12.52% CVRIDF1 and 2.32% CVRMA, above TempRMOT's best 3.78% and 0.40%.
- Adding the prediction module raises CVRIDF1 from 47.54 to 54.88 and CVRMA from 28.68 to 35.97 in-domain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CRTracker instead of the strongest reported single-view baseline | CRTrack in-domain test · All scenes | CVRIDF1: 31.45% | Helped |
| Use CRTracker instead of the strongest reported single-view baseline | CRTrack in-domain test · All scenes | CVRMA: 25.83% | Helped |
| Use CRTracker instead of the strongest reported single-view baseline | CRTrack cross-domain test · All scenes | CVRIDF1: 8.74% | Helped |
| Use CRTracker instead of the strongest reported single-view baseline | CRTrack cross-domain test · All scenes | CVRMA: 1.92% | Helped |
| Add the prediction module (ablation) | CRTracker · CRTrack in-domain test, all scenes | CVRIDF1: 47.54 -> 54.88; 7.34% | Helped |
| Add the prediction module (ablation) | CRTracker · CRTrack in-domain test, all scenes | CVRMA: 28.68 -> 35.97; 7.29% | Helped |
Surprising findings
- Cross-domain performance drops sharply for all methods, with CVRIDF1 around 2-13%, indicating poor generalization to unseen scenes and language descriptions.
- Standard CVMA can be strongly negative when false positives dominate, so the authors clip it at zero for the referring metric.
Limitations
- Performance on the cross-domain test set is very low for all methods, including CRTracker.
- Benchmark is small: 13 scenes, 344 objects and 221 language descriptions, built from DIVOTrack and CAMPUS only.
- ID switching occurs in long sequences (e.g., 6330 frames).
Experimental setup
Datasets: DIVOTrack CAMPUS Refer-KITTI
Benchmarks: CRTrack
Metrics: CVRIDF1 CVRMA CVIDF1 CVMA
Baselines: TransRMOT TempRMOT
Builds on: CrossMOT APTM MvMHAT CenterNet TransRMOT
Closest papers by meaning
- MITracker: Multi-View Integration for Visual Object Tracking (CVPR 2025 · similarity 0.86)
- HFF-Tracker: A Hierarchical Fine-grained Fusion Tracker for Referring Multi-Object Tracking (AAAI 2025 · similarity 0.85)
- iKUN: Speak to Trackers without Retraining (CVPR 2024 · similarity 0.85)
- OVTR: End-to-End Open-Vocabulary Multiple Object Tracking with Transformer (ICLR 2025 · similarity 0.84)
- Mono3DVLT: Monocular-Video-Based 3D Visual Language Tracking (CVPR 2025 · similarity 0.84)
- VastTrack: Vast Category Visual Object Tracking (NeurIPS 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.