Research map / Video / Video segmentation and tracking
UVEB: A Large-scale Benchmark and Baseline Towards Real-World Underwater Video Enhancement
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.
Builds UVEB, a 1,308-pair real-world underwater video enhancement benchmark with GT chosen via human-voted best of 20 UIE methods, and proposes UVE-Net, a supervised model that passes convolution kernels generated from a downsampled middle frame to neighboring frames.
Problem
Underwater image enhancement lacks large-scale, high-quality paired training data, and no supervised method exists for underwater video enhancement that exploits inter-frame redundancy efficiently.
Approach
Transfers the enhancement process of a low-resolution middle frame into dynamic convolution kernels that are shared with adjacent frames, enabling inter-frame information exchange at the kernel level without frame alignment or aggregation.
Key results
- UVE-Net reaches 26.27 dB PSNR and 0.4059 MSE (x10^-3) on UVEB, versus 24.21 dB for the best prior method PUIE.
- The lightweight UVE-Net-s reaches 24.43 dB and runs at about 11 FPS on 4K videos and 25 FPS on 2K videos.
- Ablation shows the auxiliary branch and FEGM/FRGM modules improve PSNR from 24.41 to 26.27 dB while lowering TFLOPs from 13.95 to 11.42.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UVE-Net instead of PUIE | UVEB video enhancement · UVEB test set | PSNR (dB): 24.21 dB to 26.27 dB | Helped |
| Use UVE-Net instead of URanker | UVEB video enhancement · UVEB test set | MSE (×103): 0.4286 to 0.4059 | Helped |
| Add AE-Net to VE-Net, with FEGM and FRGM replaced by traditional convolutions (ablation) | UVE-Net · UVEB test set | PSNR: 24.41 -> 25.15 | Helped |
| Add AE-Net to VE-Net, with FEGM and FRGM replaced by traditional convolutions (ablation) | UVE-Net · UVEB test set | MSE (×103): 0.5066 -> 0.4452 | Helped |
| Replace traditional convolutions with FEGM (ablation) | VE-Net + AE-Net · UVEB test set | PSNR: 25.15 -> 26.20 | Helped |
| Replace traditional convolutions with FEGM (ablation) | VE-Net + AE-Net · UVEB test set | MSE (×103): 0.4452 -> 0.4226 | Helped |
| Add FRGM while retaining FEGM (ablation) | VE-Net + AE-Net + FEGM · UVEB test set | PSNR: 26.20 -> 26.27 | Helped |
| Add FRGM while retaining FEGM (ablation) | VE-Net + AE-Net + FEGM · UVEB test set | MSE (×103): 0.4226 -> 0.4059 | Helped |
Surprising findings
- Some traditional and deep UIE methods (e.g., MSCNN, GDCP) produce brightness flicker across frames, while most deep learning methods do not, which led to filtering GT selections.
Limitations
- Only compared against image-level UIE methods because no supervised underwater video enhancement baselines exist.
- Ground truth is the best of 20 existing UIE outputs chosen by human observers, so GT quality is bounded by those methods.
Experimental setup
Datasets: UVEB
Benchmarks: UVEB
Metrics: PSNR MSE inference time memory usage TFLOPs
Baselines: PUIE LANet CLAHE FA+Net URanker FspiralGAN USUIR MLLE Red Channel DCP
Builds on: Charbonnier loss PixelShuffle
Closest papers by meaning
- Adaptive Dual-domain Learning for Underwater Image Enhancement (AAAI 2025 · similarity 0.89)
- Synergistic Multiscale Detail Refinement via Intrinsic Supervision for Underwater Image Enhancement (AAAI 2024 · similarity 0.87)
- Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration (CVPR 2024 · similarity 0.87)
- WaterDiffusion: Learning a Prior-involved Unrolling Diffusion for Joint Underwater Saliency Detection and Visual Restoration (AAAI 2025 · similarity 0.85)
- WebUOT-1M: Advancing Deep Underwater Object Tracking with A Million-Scale Benchmark (NeurIPS 2024 · similarity 0.85)
- Atlantis: Enabling Underwater Depth Estimation with Stable Diffusion (CVPR 2024 · similarity 0.84)
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.