Research map / Vision-language and multimodal / Human and embodied modeling

Improving Visual and Downstream Performance of Low-Light Enhancer with Vision Foundation Models Collaboration

CVPR 2025 · Vision-language and multimodal › Human and embodied modeling · approach: Vision-language pretraining · 3 citations (OpenAlex)

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 FoCo, a self-supervised low-light enhancement framework that uses CLIP, SAM and ResNet as foundation-model collaborators via feature-space adversarial alignment (FoFA) and robustness-gap pseudo-label task alignment (GaTA).

Problem

Unsupervised low-light image enhancement often lacks semantic awareness, producing results that hurt visual quality and downstream high-level task performance (classification, face detection, instance segmentation).

Approach

Combine multiple vision foundation models collaboratively: at feature level, project CLIP/SAM/ResNet features into a unified space for a discriminator (FoFA); at task level, use a robust model (CLIP) to produce voted pseudo-labels that supervise a degradation-sensitive model (ResNet) on enhanced outputs (GaTA).

Key results

• On MIT-FiveK, FoCo reaches 19.58 PSNR and 0.83 SSIM, versus 23.28 PSNR for supervised Retinexformer and 18.71 for CoLIE among unsupervised methods.
• On CODaN classification with EfficientNet, FoCo achieves 0.742 accuracy versus 0.676 for the next best unsupervised method (SGZ).
• On LIS instance segmentation, FoCo reaches 0.331 mAP-box and 0.281 mAP-seg, the best among compared methods.

Measured effects reported in the paper

Surprising findings

• Removing ResNet from FoFA degrades enhancement quality (lighting shifts) while removing CLIP or SAM hurts high-level task performance, suggesting complementary roles.
• Detection and segmentation models can substitute for classifiers in GaTA with comparable gains.

Limitations

• Pseudo-label quality depends on the robust model; very poor enhancements can yield inaccurate labels, mitigated only by voting over a memory bank.
• Gains on some tasks are modest relative to supervised baselines, e.g., Retinexformer leads on FiveK PSNR.

Experimental setup

Datasets: ExDark COCO MIT-FiveK Nikon Huawei CODaN LIS DarkFace LSRW ImageNet-1k WiderFace ADE20k

Benchmarks: MIT-FiveK Nikon Huawei CODaN DarkFace LIS

Metrics: PSNR SSIM LPIPS Top-1 Accuracy AP mAP

Baselines: Retinexformer LLFlow SKF KinD++ EnlightenGAN SCI RUAS Zero-DCE SCL-LLE SGZ CoLIE

Builds on: Zero-DCE SCI EnlightenGAN CLIP-LIT

Closest papers by meaning

• Zero-Reference Low-Light Enhancement via Physical Quadruple Priors (CVPR 2024 · similarity 0.86)
• Dispel Darkness for Better Fusion: A Controllable Visual Enhancer based on Cross-modal Conditional Adversarial Learning (CVPR 2024 · similarity 0.86)
• Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.85)
• Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.84)
• Specularity Factorization for Low-Light Enhancement (CVPR 2024 · similarity 0.84)
• Luminance-Aware Statistical Quantization: Unsupervised Hierarchical Learning for Illumination Enhancement (NeurIPS 2025 · similarity 0.84)