Research map / 3D vision / Depth and stereo
Sea-ing in Low-light
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 SelfLUID-Net, a self-supervised network that jointly recovers a restored image and depth from a single low-light underwater image, trained on the new ULVStereo stereo low-light/normal-light underwater dataset.
Problem
Joint depth estimation and image restoration from single low-light underwater (LLUW) images in real time, where existing UW methods assume normal lighting and low-light methods ignore haze and color cast; paired real training data is scarce.
Approach
Combines UW image formation and Retinex decomposition with stereo reflectance-consistency constraints (reflectance is illumination-independent, so a low-light and normal-light view share it), plus monocular video frames and a moving-pixel mask to handle occlusion and dynamic transients, all trained self-supervised on a newly captured stereo dataset.
Key results
- On UIEB dark, SelfLUID-Net reaches 17.21 dB PSNR and 0.58 SSIM, the best among compared methods.
- On FLSea depth, SelfLUID-Net achieves Pearson 0.71 and SI-MSE 0.18, the best of all compared methods.
- Runs at about 16 ms per 512x512 image (about 62 fps), providing both depth and restored image, faster than URN at 18 ms.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add photometric reprojection loss (ablation) | SelfLUID-Net · UIEB_dark | PSNR: 16.33 -> 17.21 | Helped |
| add photometric reprojection loss (ablation) | SelfLUID-Net · FLSea | Pearson correlation: 0.32 -> 0.71 | Helped |
| add stereo reflectance consistency (ablation) | SelfLUID-Net · UIEB_dark | PSNR: 16.20 -> 17.21 | Helped |
| add stereo reflectance consistency (ablation) | SelfLUID-Net · FLSea | Pearson correlation: 0.50 -> 0.71 | Helped |
| add masked monocular reprojection loss (ablation) | SelfLUID-Net · UIEB_dark | PSNR: 16.82 -> 17.21 | Helped |
| add masked monocular reprojection loss (ablation) | SelfLUID-Net · FLSea | Pearson correlation: 0.58 -> 0.71 | Helped |
| add moving-pixel mask (ablation) | SelfLUID-Net · UIEB_dark | PSNR: 16.93 -> 17.21 | Helped |
| add moving-pixel mask (ablation) | SelfLUID-Net · FLSea | Pearson correlation: 0.64 -> 0.71 | Helped |
| add spatial consistency loss (ablation) | SelfLUID-Net · UIEB_dark | PSNR: 16.50 -> 17.21 | Helped |
| add spatial consistency loss (ablation) | SelfLUID-Net · FLSea | Pearson correlation: 0.68 -> 0.71 | Helped |
Surprising findings
- Traditional UW methods HL and L2UWE obtain the highest UIQM on ULVStereo but their outputs are visually poor, showing UIQM alone is unreliable.
Limitations
- Low-light training pairs are captured by altering camera exposure, which does not exactly reproduce real low-light underwater conditions; the normal-lit image is not exact ground truth.
- Some test datasets (UIEB dark) use pseudo-GT selected subjectively, and Sea-thru GT depth is undefined over sky regions.
- Training uses only ULVStereo because no other real paired low-light/normal UW dataset exists, limiting training diversity.
Experimental setup
Datasets: ULVStereo Sea-thru NUID FLSea UIEB dark subset
Benchmarks: ULVStereo Sea-thru NUID UIEB dark FLSea
Metrics: UIQM UCIQE PSNR SSIM SI-MSE Pearson correlation
Baselines: HL HLRP UNTV MMLE CBLA USUIR URN ICSP L2UWE LIME ZeroDCE ZeroDCE++
Builds on: Monodepth2 self-supervised depth USe-ReDI-Net (URN) USUIR SCI self-calibrated illumination Retinex theory ZeroDCE
Closest papers by meaning
- Atlantis: Enabling Underwater Depth Estimation with Stable Diffusion (CVPR 2024 · similarity 0.88)
- WaterDiffusion: Learning a Prior-involved Unrolling Diffusion for Joint Underwater Saliency Detection and Visual Restoration (AAAI 2025 · similarity 0.83)
- Adaptive Dual-domain Learning for Underwater Image Enhancement (AAAI 2025 · similarity 0.83)
- UVEB: A Large-scale Benchmark and Baseline Towards Real-World Underwater Video Enhancement (CVPR 2024 · similarity 0.83)
- RestorGS: Depth-aware Gaussian Splatting for Efficient 3D Scene Restoration (CVPR 2025 · similarity 0.83)
- Diving into Underwater: Segment Anything Model Guided Underwater Salient Instance Segmentation and A Large-scale Dataset (ICML 2024 · 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.