Research map / Detection and segmentation / CNN features
Trash to Treasure: Low-Light Object Detection via Decomposition-and-Aggregation
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 T2, a low-light object detector that decomposes images into illumination and reflectance via a Retinex-based scene decomposition module and aggregates both components with a semantic (multi-scale FPN) aggregation module, without visual-quality losses.
Problem
Low-light object detection, where cascading a pre-trained image enhancer with a detector is suboptimal because enhancers optimize human-visual quality rather than detection and discard the illumination component.
Approach
Treat the Retinex-removed illumination as useful detection information rather than waste: a learned scene decomposition module (optionally initialized from an existing illumination-based enhancer) splits the input into illumination and reflectance, a weight-sharing VGG-16 extracts features from both, and a multi-scale FPN fuses them by element-wise multiplication mirroring the Retinex division, wi
Key results
- On DARK FACE, the T2 method reports about 14.8% higher mAP than HLA and 8.2% higher than the best enhancer+detector hybrid (fine-tuned SCI with DSFD), per the text.
- On ExDark with SSD, T2 reaches 59.74 mAP versus 56.85 for the best compared enhancer (ZeroDCE) and 56.30 for low-light input.
- Ablation on the proposed scene decomposition: the T2 variant (illumination plus reflectance with FPN) reaches 69.67 mAP versus 61.13 for the low-light-input baseline with no FPN (model A).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use T2 instead of ZeroDCE enhancement followed by SSD | SSD on ExDark · ExDark | mAP: 56.85 -> 59.74 | Helped |
| Use T2 instead of direct low-light input | SSD on ExDark · ExDark | mAP: 56.30 -> 59.74 | Helped |
| Use T2 with SD initialized as RUAS instead of fine-tuned RUAS enhancement | RUAS + SSD · DARK FACE | mAP (%): 65.98 -> 67.36 | Helped |
| Use T2 with SD initialized as SCI instead of fine-tuned SCI enhancement | SCI + SSD · DARK FACE | mAP (%): 65.68 -> 68.33 | Helped |
| Add joint training to RUAS enhancement and detector | RUAS + SSD · DARK FACE | mAP (%): 65.98 -> 65.70 | Hurt |
| Add joint training to SCI enhancement and detector | SCI + SSD · DARK FACE | mAP (%): 65.68 -> 65.42 | Hurt |
| Add FPN to low-light image input (ablation) | A (low-light input) · DARK FACE | mAP (%): 61.13 -> 66.42 | Helped |
| Add FPN to reflectance input (ablation) | D (reflectance input) · DARK FACE | mAP (%): 66.94 -> 67.68 | Helped |
| Use illumination instead of low-light image as the single input (ablation) | Single-component detector without FPN · DARK FACE | mAP (%): 61.13 -> 59.28 | Hurt |
| Use reflectance instead of low-light image as the single input (ablation) | Single-component detector without FPN · DARK FACE | mAP (%): 61.13 -> 66.94 | Helped |
| Add illumination alongside reflectance and aggregate the components with FPN (ablation) | Reflectance-input detector with FPN · DARK FACE | mAP (%): 67.68 -> 69.67 | Helped |
| Add FPN to illumination input (ablation) | C (illumination input) · DARK FACE | mAP (%) | No clear effect |
Surprising findings
- Classical enhancer MSRCR with poor visual quality was reported to work better for detection than more advanced enhancers.
- Using only the illumination component (59.28 mAP) performs worse than low-light input with FPN, while reflectance alone reaches 66.94 mAP.
Limitations
- Experiments mostly on DARK FACE face detection; generality to general object detection is only shown on ExDark.
- Code is promised only upon acceptance, with no release link.
- Hardware and training compute are not stated.
Experimental setup
Datasets: DARK FACE ExDark
Benchmarks: DARK FACE ExDark
Metrics: mAP precision-recall curves
Baselines: RUAS SCI DeepUPE DRBN EnGAN KinD ZeroDCE HLA REG
Builds on: Retinex theory feature pyramid network DSFD S3FD PyramidBox
Closest papers by meaning
- Self-Guided Low Light Object Detection Framework (ICLR 2026 · similarity 0.87)
- Boosting Object Detection with Zero-Shot Day-Night Domain Adaptation (CVPR 2024 · similarity 0.85)
- You Only Look Around: Learning Illumination-Invariant Feature for Low-light Object Detection (NeurIPS 2024 · similarity 0.85)
- Not All Tokens Matter All The Time: Dynamic Token Aggregation Towards Efficient Detection Transformers (ICML 2025 · similarity 0.85)
- DoGA: Enhancing Grounded Object Detection via Grouped Pre-Training with Attributes (AAAI 2025 · similarity 0.85)
- Roboflow100-VL: A Multi-Domain Object Detection Benchmark for Vision-Language Models (NeurIPS 2025 · 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.