Research map / Detection and segmentation / Open-set and few-shot detection
Boosting Object Detection with Zero-Shot Day-Night Domain Adaptation
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 DAI-Net, a zero-shot day-night domain adaptation detector that learns Retinex-based illumination-invariant reflectance features from well-lit images and synthetic low-light pairs, evaluated on DARK FACE, ExDark and CODaN.
Problem
Object detectors trained on well-lit images degrade badly on low-light images, and collecting and annotating real low-light data is difficult, so the paper targets generalizing detectors from well-lit to low-light scenes without any low-light target data.
Approach
Adds a reflectance decoder branch to an established detector, supervised by pseudo reflectance from a pretrained Retinex decomposition net and a mutual feature alignment loss, plus an interchange-redecomposition-coherence procedure with a redecomposition cohering loss that enforces consistency across two sequential Retinex decompositions.
Key results
- On WIDER FACE→DARK FACE with DSFD, DAI-Net reaches 28.0% mAP versus 25.7% for the prior zero-shot method Sim-MinMax.
- On COCO→ExDark with YOLOv3, DAI-Net achieves 78.3 total mAP versus 74.0 for MAET.
- On CODaN classification, DAI-Net reaches 68.44% top-1 versus 65.87% for Sim-MinMax.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DAI-Net instead of Sim-MinMax | DSFD · WIDER FACE→DARK FACE test set | mAP(%): 25.7 -> 28.0 | Helped |
| Fine-tune DAI-Net on labeled DARK FACE training data | DSFD · WIDER FACE→DARK FACE test set | mAP(%): 46.0 -> 52.9 | Helped |
| Use DAI-Net instead of MAET | YOLOv3 · COCO→DARK FACE validation set | mAP(%): 55.8 -> 57.0 | Helped |
| Use DAI-Net instead of MAET | YOLOv3 · ExDark | mAP(%): 74.0 -> 78.3 | Helped |
| Use DAI-Net instead of Sim-MinMax | ResNet-18 · CODaN | Top-1 (%): 65.87 -> 68.44 | Helped |
| Train directly on synthesized low-light images (ablation) | DSFD · WIDER FACE→DARK FACE | mAP(%): 15.2% -> 11.6% | Hurt |
| Add a reflectance decoding branch (ablation) | DSFD · WIDER FACE→DARK FACE | mAP(%): 15.2% -> 20.5% | Helped |
| Decode illumination instead of reflectance (ablation) | DSFD with an auxiliary decoding branch · WIDER FACE→DARK FACE | mAP(%): 20.5 -> 16.5 | Hurt |
| Decode both reflectance and illumination (ablation) | DSFD with an auxiliary decoding branch · WIDER FACE→DARK FACE | mAP(%): 20.5 -> 19.0 | Hurt |
| Use redecomposition cohering loss instead of no decomposition process (ablation) | DAI-Net with reflectance decoding · WIDER FACE→DARK FACE | mAP(%): 20.5% -> 22.3% | Helped |
| Use both penalty loss and redecomposition cohering loss instead of redecomposition cohering loss alone (ablation) | DAI-Net with reflectance decoding · WIDER FACE→DARK FACE | mAP(%): 22.3 -> 22.0 | No clear effect |
| Apply three sequential decompositions instead of two (ablation) | DAI-Net · WIDER FACE→DARK FACE | mAP(%) | No clear effect |
Surprising findings
- Training directly on Dark ISP-synthesized low-light images degrades DSFD from 15.2% to 11.6% mAP on DARK FACE, showing synthetic low-light images differ from real ones.
Limitations
- Learning decomposition on the fly with detection was unstable, so a pretrained Retinex net is needed for pseudo labels.
- Synthetic low-light images from Dark ISP are not equivalent to real low-light data, so direct training on them hurts performance.
Experimental setup
Datasets: WIDER FACE COCO 2017 DARK FACE ExDark CODaN LOL-v2
Benchmarks: DARK FACE ExDark CODaN
Metrics: mAP top-1 accuracy
Baselines: Faster-RCNN SSH RetinaFace SRN SFA ParamidBox SmallHardFace DSFD CIConv Sim-MinMax HLA-Face MAET
Builds on: DSFD YOLOv3 RetinexNet MAET Sim-MinMax
Closest papers by meaning
- FRBNet: Revisiting Low-Light Vision through Frequency-Domain Radial Basis Network (NeurIPS 2025 · similarity 0.87)
- You Only Look Around: Learning Illumination-Invariant Feature for Low-light Object Detection (NeurIPS 2024 · similarity 0.86)
- ISP-Teacher:Image Signal Process with Disentanglement Regularization for Unsupervised Domain Adaptive Dark Object Detection (AAAI 2024 · similarity 0.85)
- Trash to Treasure: Low-Light Object Detection via Decomposition-and-Aggregation (AAAI 2024 · similarity 0.85)
- NightCC: Nighttime Color Constancy via Adaptive Channel Masking (CVPR 2024 · similarity 0.85)
- Zero-Reference Low-Light Enhancement via Physical Quadruple Priors (CVPR 2024 · similarity 0.85)
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.