Research map / Detection and segmentation / CNN features
You Only Look Around: Learning Illumination-Invariant Feature for Low-light Object Detection
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 YOLA, an Illumination-Invariant Module that learns zero-mean convolutional kernels over log cross-channel color ratios (Lambertian-motivated) to produce illumination-invariant features that plug into detectors for low-light object detection.
Problem
Object detection in low-light conditions suffers from illumination effects that degrade features; prior work relies on image enhancement, synthetic data, or task-specific losses.
Approach
Derive illumination-invariant features from the Lambertian model using cross-color ratios of neighboring pixels, then generalize the fixed kernel to learnable convolutional kernels constrained to zero mean (plus an illumination-invariant consistency loss), trained end-to-end with the detector.
Key results
- On ExDark with TOOD, YOLA reaches 75.2 mAP50 vs. 72.5 baseline and 74.3 for MAET; on UG2+ DARK FACE with TOOD it reaches 67.4 vs. 62.1 baseline.
- With YOLOv3 on ExDark, YOLA improves mAP50 from 71.0 to 72.7 with only 0.008M parameters.
- On COCO 2017 val, TOOD+YOLA improves mAP from 40.5 to 41.2 in the over-light setting and from 41.7 to 42.3 in the well-lit setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add illumination-invariant features with IIM (ablation) | TOOD · ExDark | mAP50: 72.5 -> 74.7 | Helped |
| Add illumination-invariant features with IIM (ablation) | TOOD · DarkFace | mAP50: 62.1 -> 66.9 | Helped |
| Use fixed edge kernels instead of no illumination module (ablation) | TOOD · ExDark | mAP50: 72.5 -> 73.8 | Helped |
| Use fixed edge kernels instead of no illumination module (ablation) | TOOD · DarkFace | mAP50: 62.1 -> 64.5 | Helped |
| Replace fixed edge kernels with learnable kernels (ablation) | TOOD with IIM · ExDark | mAP50: 73.8 -> 74.7 | Helped |
| Replace fixed edge kernels with learnable kernels (ablation) | TOOD with IIM · DarkFace | mAP50: 64.5 -> 66.9 | Helped |
| Add the zero-mean constraint to learnable kernels (ablation) | TOOD with IIM · ExDark | mAP50: 74.7 -> 75.2 | Helped |
| Add the zero-mean constraint to learnable kernels (ablation) | TOOD with IIM · DarkFace | mAP50: 66.9 -> 67.4 | Helped |
| Add YOLA | TOOD · COCO 2017val well-lit | mAP: 41.7 -> 42.3 | Helped |
| Add YOLA | TOOD · COCO 2017val over-light | mAP: 40.5 -> 41.2 | Helped |
| Add YOLA | YOLOv3 · ExDark | mAP50: 71.0 -> 72.7 | Helped |
| Add YOLA | TOOD · ExDark | mAP50: 72.5 -> 75.2 | Helped |
Surprising findings
- Naive YOLA (MSE-consistency without IIM) underperforms the baseline on TOOD for ExDark (71.6 vs 72.5 mAP50).
- Several LLIE methods such as SMG and NeRCo degrade detection performance relative to the un-enhanced baseline on UG2+ DARK FACE.
Limitations
- Relies on the Lambertian assumption and the approximation that illumination is locally uniform across neighboring pixels.
- Zero-mean constraint alone can allow trivial kernels with distant-pixel cancellation, requiring an additional consistency loss.
Experimental setup
Datasets: ExDark UG2+ DARK FACE COCO 2017 LIS
Benchmarks: ExDark UG2+ DARK FACE COCO 2017 val (well-lit and over-lit) LIS
Metrics: mAP50 recall mAP AP50 AP75 model size
Baselines: KIND SMG NeRCo DENet GDIP IAT MAET YOLA-Naive FeatEnHancer Zero-DCE
Builds on: cross color ratio YOLOv3 TOOD MAET
Closest papers by meaning
- FRBNet: Revisiting Low-Light Vision through Frequency-Domain Radial Basis Network (NeurIPS 2025 · similarity 0.87)
- Boosting Object Detection with Zero-Shot Day-Night Domain Adaptation (CVPR 2024 · similarity 0.86)
- A New Multi-Source Light Detection Benchmark and Semi-Supervised Focal Light Detection (NeurIPS 2024 · similarity 0.85)
- Trash to Treasure: Low-Light Object Detection via Decomposition-and-Aggregation (AAAI 2024 · similarity 0.85)
- Alignment-Free RGB-T Salient Object Detection: A Large-Scale Dataset and Progressive Correlation Network (AAAI 2025 · similarity 0.84)
- DiMSOD: A Diffusion-Based Framework for Multi-Modal Salient Object Detection (AAAI 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.