Research map / Detection and segmentation / Tracking
Omnidirectional 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.
Proposes OmniTrack, a panoramic multi-object tracking framework unifying tracking-by-detection and end-to-end paradigms with a feedback mechanism, and introduces the QuadTrack panoramic MOT dataset collected from a quadruped robot.
Problem
Multi-object tracking in 360° panoramic imagery is underexplored; pinhole-based MOT methods degrade under panoramic distortions (resolution loss, geometric deformation, uneven lighting) and rapid sensor motion.
Approach
Feedback mechanism that feeds trajectory information (tracklets) back into the detector via FlexiTrack Instances, plus a CircularStatE Module (dynamic convolution and state space model) to mitigate panoramic distortion; the same framework yields an E2E tracker (OmniTrackE2E), a TBD tracker (OmniTrackDA), or a detector (OmniTrackDet).
Key results
- OmniTrackDA achieves 26.92% HOTA on the JRDB test set, a 3.43-point gain over prior methods (HybridSORT at 25.01%).
- OmniTrackDA reaches 23.45% HOTA on the QuadTrack test set, surpassing the baseline by 6.81%.
- OmniTrackE2E reaches 21.56% HOTA on JRDB test, beating MOTRv2 by 3.34 points, and trains over four times faster than other E2E methods (20 vs 80-130 hours).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OmniTrackDA instead of OC-SORT | JRDB tracking · JRDB test | HOTA: 25.04 -> 26.92 | Helped |
| Use OmniTrackDA instead of DeepSORT | QuadTrack tracking · QuadTrack test | HOTA: 21.16 -> 23.45 | Helped |
| Use OmniTrackDA detection and tracking instead of YOLO11 with HybridSORT | JRDB tracking · JRDB val | HOTA: 29.71 -> 31.05 | Helped |
| Add denoise instances (ablation) | OmniTrackE2E · JRDB val | HOTA: 3.80 -> 24.32 | Helped |
| Add FlexiTrack Instances (ablation) | OmniTrackE2E with denoise instances · JRDB val | HOTA: 24.32 -> 25.12 | Helped |
| Add FlexiTrack Instances (ablation) | OmniTrackE2E without denoise instances · JRDB val | HOTA: 0.01 -> 3.80 | Helped |
| Replace DynamicSSM with MLP on S3, S4, and S5 (ablation) | CircularStatE · JRDB val | HOTA: 25.120 -> 21.951 | Hurt |
| Replace DynamicSSM with Conv on S3, S4, and S5 (ablation) | CircularStatE · JRDB val | HOTA: 25.120 -> 23.565 | Hurt |
| Add DynamicSSM to S5 (ablation) | CircularStatE · JRDB val | HOTA: 23.296 -> 24.426 | Helped |
| Add DynamicSSM to S4 (ablation) | CircularStatE · JRDB val | HOTA: 23.296 -> 24.539 | Helped |
| Add DynamicSSM to S3 (ablation) | CircularStatE · JRDB val | HOTA: 23.296 -> 25.120 | Helped |
Surprising findings
- OmniTrackDet improves detection-based tracking only marginally in HOTA (about 0.2% on average) while lowering OSPA and raising MOTA, suggesting detector gains alone are limited.
Limitations
- Limited to 2D panoramic tracking without depth or 3D capabilities.
- Method is centered on a mobile robotic platform and may not generalize beyond it.
- OmniTrackE2E has negative MOTA on QuadTrack, and OmniTrackDA is slower than the YOLO11 baseline (about 11.8 vs 49 FPS).
Experimental setup
Datasets: JRDB QuadTrack
Benchmarks: JRDB QuadTrack
Metrics: HOTA IDF1 OSPA MOTA DetA AssA FPS
Baselines: TrackFormer MOTRv2 MOTR SORT DeepSORT ByteTrack Bot-SORT OC-SORT HybridSORT DiffMOT
Builds on: MOTR MOTRv2 OC-SORT ByteTrack RT-DETR
Closest papers by meaning
- More Than Meets the Eye: Enhancing Multi-Object Tracking Even with Prolonged Occlusions (ICML 2025 · similarity 0.85)
- NetTrack: Tracking Highly Dynamic Objects with a Net (CVPR 2024 · similarity 0.85)
- UCMCTrack: Multi-Object Tracking with Uniform Camera Motion Compensation (AAAI 2024 · similarity 0.85)
- M3SOT: Multi-Frame, Multi-Field, Multi-Space 3D Single Object Tracking (AAAI 2024 · similarity 0.84)
- TAPVid-360: Tracking Any Point in 360 from Narrow Field of View Video (NeurIPS 2025 · similarity 0.83)
- Collaborating Vision, Depth, and Thermal Signals for Multi-Modal Tracking: Dataset and Algorithm (NeurIPS 2025 · similarity 0.83)
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.