Research map / Detection and segmentation / Tracking
Multi-Object Tracking in the Dark
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.
Builds LMOT, a dual-camera RAW low-light multi-object tracking dataset with aligned well-lit videos, and proposes LTrack, which adds an adaptive low-pass downsample module and a degradation suppression learning loss to a YOLOX-based tracker.
Problem
Multi-object tracking in low-light scenes is rarely studied; low-light sensor noise degrades detectors and appearance features, and no MOT dataset exists for dark scenes.
Approach
Use a dual-camera rig to capture pixel-aligned low-light and well-lit RAW video pairs, annotate on the well-lit video, and learn noise-invariant features directly from low-light input (via spatial low-pass downsampling and a feature-matching loss to well-lit features) instead of cascading enhancement.
Key results
- LTrack with OC-SORT reaches 35.1 HOTA on LMOT-real versus 32.2 for the RAOD baseline and 34.6 for Base-all (Table 6).
- On the LMOT validation set with ByteTrack, LTrack improves HOTA from 28.1 (Base-all) to 29.4 while keeping roughly the same parameters and FLOPs.
- Degradation suppression learning gives about a 1-point HOTA gain over the baseline in the ablation (28.1 to 29.0), and combining it with ALD yields 29.4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add degradation suppression learning (DSL) without ALD (ablation) | LTrack · LMOT test | HOTA: 28.1 -> 29.0 | Helped |
| Add adaptive low-pass downsampling (ALD) without DSL (ablation) | LTrack · LMOT test | HOTA: 28.1 -> 28.3 | Helped |
| Add both ALD and DSL (ablation) | LTrack · LMOT test | HOTA: 28.1 -> 29.4 | Helped |
| Use sRGB instead of RAW 12-bit input (ablation) | LTrack · LMOT test | HOTA: 29.4 -> 29.2 | Hurt |
| Use RAW 8-bit instead of RAW 12-bit input (ablation) | LTrack · LMOT test | HOTA: 29.4 -> 29.0 | Hurt |
| Use RAW 10-bit instead of RAW 12-bit input (ablation) | LTrack · LMOT test | HOTA: 29.4 -> 29.3 | Hurt |
| Use LTrack instead of Base-all | OC-SORT on LMOT-real · LMOT-real | HOTA: 34.6 -> 35.1 | Helped |
| Use LTrack instead of SDSD | ByteTrack on LMOT test · LMOT test | HOTA: 29.1 -> 29.4 | Helped |
| Use LTrack instead of LLFlow | OC-SORT on LMOT test · LMOT test | HOTA: 29.4 -> 29.8 | Helped |
Surprising findings
- A detector trained only on well-lit images drops from 37.0 mAP on well-lit to 3.1 mAP on low-light images.
- Using only appearance matching for association with low-light inputs performs very poorly (HOTA 53.8 vs 86.1 under well-lit), indicating appearance cues collapse in the dark.
Limitations
- Only considers low-light degradation; other adverse conditions such as rain and fog are not addressed.
- Truck and bicycle tracking performance is low due to few training instances.
- Dataset is restricted to city outdoor scenes and six object categories.
Experimental setup
Datasets: LMOT COCO MOT17 KITTI DanceTrack SportMOT BDD100K
Benchmarks: LMOT test set LMOT-real
Metrics: HOTA AssA IDF1 MOTA DetA mAP AP50 AP75
Baselines: ByteTrack OC-SORT LLFlow SDSD DNF SMOID RAOD Base-low Base-well Base-all
Builds on: ByteTrack OC-SORT YOLOX Total Variation loss
Closest papers by meaning
- All-Day Multi-Camera Multi-Target Tracking (CVPR 2025 · similarity 0.85)
- Sampling-Resilient Multi-Object Tracking (AAAI 2024 · similarity 0.83)
- Towards Robust Event-guided Low-Light Image Enhancement: A Large-Scale Real-World Event-Image Dataset and Novel Approach (CVPR 2024 · similarity 0.83)
- DiffusionTrack: Diffusion Model for Multi-Object Tracking (AAAI 2024 · similarity 0.83)
- Boosting Object Detection with Zero-Shot Day-Night Domain Adaptation (CVPR 2024 · similarity 0.82)
- UCMCTrack: Multi-Object Tracking with Uniform Camera Motion Compensation (AAAI 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.