Research map / Label-efficient and robust learning / Domain adaptation and generalization
Towards Generalized Representations for Low-Light Understanding: When Signal Constancy Meets Semantic Enrichment
Yifan Li, Haofeng Huang, Wenhan Yang, Jiaying Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add illumination-invariant prior and VFM semantic enrichment (UniPrior) without real low-light data | ResNet-18 · CODaN classification | Top-1 accuracy (%): 53.32 -> 74.52 | Improved |
| add illumination-invariant prior as auxiliary input (ablation) | UniPrior · CODaN classification | Top-1 accuracy (%): 67.24 -> 74.52 | Improved |
| zero-initialize the prior channels in the merge convolution (ablation) | UniPrior · CODaN classification | Top-1 accuracy (%): 62.25 -> 74.52 | Improved |
| add decoding-based regularization (prior decoder) (ablation) | UniPrior · CODaN classification | Top-1 accuracy (%): 70.26 -> 74.52 | Improved |
| add VFM contrastive augmentation (ablation) | UniPrior · CODaN classification | Top-1 accuracy (%): 68.89 -> 74.52 | Improved |
| use correlation-map alignment instead of naive L1 feature alignment (ablation) | UniPrior · CODaN classification | Top-1 accuracy (%): 63.26 -> 74.52 | Improved |
| use outlier-robust loss for prior consistency instead of plain L1 (ablation) | UniPrior · CODaN classification | Top-1 accuracy (%): 73.63 -> 74.52 | Improved |
| add test-time adaptive machine-oriented enhancer | UniPrior · CODaN classification | Top-1 accuracy (%): 74.52 -> 75.72 | Improved |
| add the TTA enhancer (inference time cost) | UniPrior · CODaN classification | inference time (s): 0.005 -> 2.593 | Hurt |
| add CLIP-based intermediate loss to enhancer (ablation) | UniPrior TTA enhancer · CODaN classification | Top-1 accuracy (%): 74.86 -> 75.72 | No clear effect |
| human-oriented low-light enhancement before understanding | RefineNet · Nighttime Driving segmentation | mIoU (%): 34.3 -> 25.2 | Hurt |
| change low-light training distribution for data-requiring adaptation | WiiD · CODaN classification | accuracy drop (%): 1.72% drop | Hurt |
Closest papers by meaning
- From Enhancement to Understanding: Build a Generalized Bridge for Low-light Vision via Semantically Consistent Unsupervised Fine-tuning (ICCV 2025 · similarity 0.90)
- Unsupervised Illumination Adaptation for Low-Light Vision (TPAMI 2024 · similarity 0.88)
- Semantically Contrastive Learning for Low-Light Image Enhancement (AAAI 2022 · similarity 0.88)
- Zero-Reference Low-Light Enhancement via Physical Quadruple Priors (CVPR 2024 · similarity 0.87)
- See through the Dark: Learning Illumination-affined Representations for Nighttime Occupancy Prediction (NeurIPS 2025 · similarity 0.85)
- Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · 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.
