Research map / Detection and segmentation / Tracking
MM-Tracker: Motion Mamba for UAV-platform Multiple 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 MM-Tracker, a UAV multiple object tracker with a Motion Mamba module for global-plus-local motion estimation from bi-temporal detection features and a Motion Margin loss that raises decision boundaries for fast-moving (motion-blurred) objects.
Problem
Multiple object tracking from UAV platforms faces local object motion plus global camera motion, and motion blur from large motion makes detection of fast-moving objects hard; prior motion models are local-only or ignore blur, and long-tailed motion distributions under-train hard samples.
Approach
Estimate object motion from detector features (not raw images) by combining local cross-correlation of consecutive frame features with bidirectional Mamba scans for global context, and add a motion-dependent classification margin (MMLoss) so high-motion objects receive larger decision boundaries.
Key results
- MM-Tracker reaches 44.7 MOTA and 58.3 IDF1 on VisDrone and 51.4 MOTA and 68.9 IDF1 on UAVDT, the best among compared trackers.
- Motion Mamba module runs in 6.9 ms versus 115 ms for EMD-Flow, with MOTA 40.9 vs 40.7 on VisDrone.
- MMLoss improves MOTA from 39.6 to 42.1 on VisDrone over standard cross-entropy; Focal loss drops to 36.4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Motion Margin Loss (ablation) | MM-Tracker baseline · Visdrone test set | MOTA: 39.6 -> 42.1 | Helped |
| Add Motion Margin Loss (ablation) | MM-Tracker baseline · UAVDT test set | MOTA: 47.5 -> 49.2 | Helped |
| Add Motion Mamba (ablation) | MM-Tracker baseline · Visdrone test set | MOTA: 39.6 -> 40.9 | Helped |
| Add Motion Mamba (ablation) | MM-Tracker baseline · UAVDT test set | MOTA: 47.5 -> 49.7 | Helped |
| Combine Motion Margin Loss and Motion Mamba (ablation) | MM-Tracker baseline · Visdrone test set | MOTA: 39.6 -> 44.7 | Helped |
| Combine Motion Margin Loss and Motion Mamba (ablation) | MM-Tracker baseline · UAVDT test set | MOTA: 47.5 -> 51.4 | Helped |
| Use local correlation and bidirectional SSM instead of the baseline motion modeling (ablation) | MM-Tracker · Visdrone test set | MOTA: 39.6 -> 40.9 | Helped |
| Replace Motion Mamba with FastFlowNet motion prediction (ablation) | MM-Tracker · Visdrone test set | MOTA: 40.9 -> 40.4 | Hurt |
| Replace Motion Mamba with GMA motion prediction (ablation) | MM-Tracker · Visdrone test set | MOTA: 40.9 -> 40.5 | Hurt |
| Replace Motion Mamba with EMD-Flow motion prediction (ablation) | MM-Tracker · Visdrone test set | MOTA: 40.9 -> 40.7 | Hurt |
| Replace Motion Mamba with EMD-Flow motion prediction (ablation) | MM-Tracker · Visdrone test set | inference time (ms): 6.9 vs 115 | Hurt |
| Replace MMLoss with Focal loss (ablation) | MM-Tracker detector · Visdrone test-dev | MOTA (%): 42.1 -> 36.4 | Hurt |
Surprising findings
- Focal loss performed worse than cross-entropy (36.4 vs 39.6 MOTA) on VisDrone, while the proposed margin loss helped.
Limitations
- Motion ground truth depends on a pretrained optical flow estimator (EMD-Flow) and annotated object offsets.
- Evaluated only on two UAV tracking datasets (VisDrone, UAVDT) with a single YOLOX-S detector.
Experimental setup
Datasets: VisDrone UAVDT
Benchmarks: VisDrone UAVDT
Metrics: MOTA IDF1 FPS inference time (ms)
Baselines: SiamMOT FairMOT ByteTrack UAVMOT OC-SORT FOLT U2MOT TrackSSM Focal loss LDAM
Builds on: Mamba YOLOX OC-SORT FOLT EMD-Flow
Closest papers by meaning
- MMOT: The First Challenging Benchmark for Drone-based Multispectral Multi-Object Tracking (NeurIPS 2025 · similarity 0.86)
- MUST: The First Dataset and Unified Framework for Multispectral UAV Single Object Tracking (CVPR 2025 · similarity 0.84)
- UCMCTrack: Multi-Object Tracking with Uniform Camera Motion Compensation (AAAI 2024 · similarity 0.83)
- More Than Meets the Eye: Enhancing Multi-Object Tracking Even with Prolonged Occlusions (ICML 2025 · similarity 0.83)
- M$^2$E-UAV: A Benchmark and Analysis for Onboard Motion-on-Motion Event-Based Tiny UAV Detection (NeurIPS 2026 · similarity 0.83)
- UAV3D: A Large-scale 3D Perception Benchmark for Unmanned Aerial Vehicles (NeurIPS 2024 · 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.