Research map / 3D vision / Depth and stereo
All-Day Multi-Camera Multi-Target 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 M3Track, the first RGB-thermal multi-camera multi-target tracking dataset with low-light sequences, and ADMCMT, a DLA-34/CenterNet + GMT tracker with Mamba-based RGB-IR fusion guided by a lighting classifier and a nearby-target re-ID strategy.
Problem
Multi-camera multi-target tracking methods are designed for daytime with good lighting and fail under low-light (nighttime) conditions due to lack of visible appearance features.
Approach
Add infrared (thermal) modality to MCMT tracking, fuse RGB and IR features with a Mamba-based module whose fusion is guided by a lighting-relevant feature predicted from the visible branch, and enrich target re-ID features with nearby-object features collected from overlapping boxes.
Key results
- On M3Track, ADMCMT reaches 70.94 MOTA and 71.47 IDF1 single-camera, beating the second-best method by 2.73 MOTA and 9.45 IDF1.
- Multi-camera CVMA reaches 61.81 and CVIDF1 68.77, improvements of 16.79 CVMA and 13.96 CVIDF1 over the best baseline.
- On night sequences, ADMCMT outperforms the second-best model by 2.88 MOTA and 8.6 CVIDF1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| introduce ADMCMT with RGB and infrared information | MCMT tracking · M3Track All | CVMA: 45.02 -> 61.81 | Helped |
| introduce ADMCMT with RGB and infrared information | MCMT tracking · M3Track Night | CVIDF1: 49.45 -> 58.05 | Helped |
| add the ADMF fusion channels to simple averaged RGB-infrared fusion (ablation) | ADMCMT · M3Track | MOTA: 58.98 -> 68.23 | Helped |
| add the Lighting Guidance Model (ablation) | ADMCMT ADMF · M3Track | MOTA: 68.62 -> 69.16 | Helped |
| add the LI channel (ablation) | ADMCMT ADMF · M3Track | MOTA: 69.16 -> 70.47 | Helped |
| add the LGM and LI channel (ablation) | ADMCMT ADMF · Day | MOTA: 69.55 -> 71.75 | Helped |
| add the LGM and LI channel (ablation) | ADMCMT ADMF · Night | CVIDF1: 54.30 -> 58.52 | Helped |
| add 1 nearby target (ablation) | ADMCMT · M3Track | CVIDF1: 67.09 -> 67.86 | Helped |
| increase nearby targets from 1 to 3 (ablation) | ADMCMT · M3Track | CVIDF1: 67.86 -> 51.46 | Hurt |
| set area threshold to 0.6 (ablation) | ADMCMT Nearby Target Collection · M3Track | MOTA: 70.63, 70.61, 70.94, 70.51, 70.71 | Mixed |
Surprising findings
- Simple pixel-level averaging of RGB and IR features degraded MDA and MOTP relative to the single visible baseline, suggesting naive fusion adds noise.
Limitations
- Evaluation limited to a single new dataset (M3Track) with 88 sequences from two drones.
- Nearby-target count and area threshold must be tuned depending on camera pitch and object density.
Experimental setup
Datasets: M3Track
Benchmarks: M3Track
Metrics: MOTA HOTA IDF1 DetA MOTP MDA CVIDF1 CVMA
Baselines: OSNet CrossMOT CT MvMHAT AGW
Builds on: CenterNet GMT VSS block (VMamba) DLA-34
Closest papers by meaning
- Multi-Object Tracking in the Dark (CVPR 2024 · similarity 0.85)
- MTMMC: A Large-Scale Real-World Multi-Modal Camera Tracking Benchmark (CVPR 2024 · similarity 0.85)
- Collaborating Vision, Depth, and Thermal Signals for Multi-Modal Tracking: Dataset and Algorithm (NeurIPS 2025 · similarity 0.85)
- Bi-directional Adapter for Multimodal Tracking (AAAI 2024 · similarity 0.84)
- UCMCTrack: Multi-Object Tracking with Uniform Camera Motion Compensation (AAAI 2024 · similarity 0.83)
- Single-Model and Any-Modality for Video Object Tracking (CVPR 2024 · similarity 0.82)
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.