Research map / Video / Video segmentation and tracking
LiBrA-Net: Lie-Algebraic Bilateral Affine Fields for Real-Time 4K Video Dehazing
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 LiBrA-Net, a Lie-algebraic bilateral-grid video dehazer that predicts low-resolution spatial-color and temporal affine sub-grids, fuses them in gl(3) and maps them to GL(3) via a Cayley map, plus a light high-frequency refiner, and releases the UHV-4K synthetic 4K video dehazing benchmark.
Problem
Video dehazing at native ultra-high-definition (4K) resolution is hard: existing video dehazers are designed for 1080p or below and their cost scales with pixel count, while no paired 4K video dehazing benchmark exists.
Approach
Factorize the spatiotemporal affine color transform implied by the atmospheric scattering model into two low-resolution bilateral sub-grids (spatial-color and temporal) whose coefficients are added in the gl(3) Lie algebra and mapped to invertible GL(3) transforms via a Cayley map, so that prediction cost is independent of output resolution and only trilinear slicing scales with pixels.
Key results
- On UHV-4K, LiBrA-Net reaches 24.28 dB PSNR, 0.9437 SSIM, and 0.0401 LPIPS, the best among compared methods.
- At native 4K, LiBrA-Net runs at 25.27 FPS with 0.28K GFLOPs and 6.12M parameters on a single V100 32GB.
- Removing the HF-Refiner drops UHV-4K PSNR by 3.41 dB; removing the chromatic or temporal field drops it by about 2.2 dB.
Surprising findings
- Removing the temporal Lie regularizer lowers tOF by 7.8% while per-video sigma-PSNR still rises by 44%, so tOF alone is not a reliable steadiness measure here.
- Compared methods with the lowest tOF (DCL, ViWS-Net) trail LiBrA-Net in PSNR by 3.47 dB and 2.22 dB, suggesting low tOF partly reflects over-smoothing.
Limitations
- Temporal window covers roughly one second (T=5 frames); minute-scale coherence is not modeled.
- Assumes a single beta and A_infinity per video, so mixed-weather conditions are out of scope.
- UHV-4K is synthetic; paired real 4K hazy/clean data does not exist, and haze realism trails real-capture datasets.
Experimental setup
Datasets: UHV-4K REVIDE HazeWorld Inter4K UVG
Benchmarks: UHV-4K REVIDE HazeWorld
Metrics: PSNR SSIM LPIPS tOF sigma-PSNR FPS GFLOPs Params
Baselines: MSBDN-DFF FFA-Net 4KDehazing DeHamer DehazeFormer UHDformer UHDDIP DehazeXL CG-IDN PM-Net MAP-Net ViWS-Net
Builds on: 4KDehazing HDRNet bilateral grid Bilateral Guided Upsampling DINOv2
Closest papers by meaning
- Depth-Centric Dehazing and Depth-Estimation from Real-World Hazy Driving Video (AAAI 2025 · similarity 0.87)
- 4DGC: Rate-Aware 4D Gaussian Compression for Efficient Streamable Free-Viewpoint Video (CVPR 2025 · similarity 0.85)
- Exploiting Deblurring Networks for Radiance Fields (CVPR 2025 · similarity 0.84)
- DNGaussian: Optimizing Sparse-View 3D Gaussian Radiance Fields with Global-Local Depth Normalization (CVPR 2024 · similarity 0.84)
- 4K4D: Real-Time 4D View Synthesis at 4K Resolution (CVPR 2024 · similarity 0.84)
- Tokenize Image Patches: Global Context Fusion for Effective Haze Removal in Large Images (CVPR 2025 · 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.