Research map / Video / Video segmentation and tracking
Depth-Centric Dehazing and Depth-Estimation from Real-World Hazy Driving Video
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 Depth-Centric Learning (DCL), which couples the atmospheric scattering model and a brightness consistency constraint through a shared depth network, plus two misaligned-reference discriminators, to jointly dehaze and estimate depth from real hazy driving video.
Problem
Jointly removing haze and estimating depth from real-world monocular hazy driving video, where paired ground truth is hard to obtain and existing ASM-based dehazing and self-supervised depth methods degrade in real fog.
Approach
A unified ASM-BCC model in which one shared depth network is supervised both by atmospheric scattering reconstruction (using dehazed adjacent frames) and by brightness-consistency photometric reprojection, combined with misaligned clear-video regularization via two discriminators (MFIR for wavelet high-frequency detail, MDR for depth black holes).
Key results
- DCL achieves FADE 0.6914 and NIQE 3.4412 on GoProHazy, ranking first across the three dehazing datasets and beating DVD's NIQE by 22.04%.
- On DENSE-Fog (light), DCL reaches abs Rel 0.311 and RMSE log 0.364, improving over the best compared self-supervised depth baseline (RobustDepth, 0.316 and 0.370).
- DCL runs at 0.075 s per frame at 640x192 with 11.38M params, the fastest among compared dehazing methods, since only the dehazing subnetwork is used at test time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DCL instead of DVD | Video dehazing · GoProHazy | FADE: 0.7061 -> 0.6914 | Helped |
| use DCL instead of DVD | Video dehazing · DrivingHazy | NIQE: 4.4820 -> 3.5329 | Helped |
| use DCL instead of NSDNet | Video dehazing · InternetHazy | NIQE: 4.3144 -> 3.5545 | Helped |
| use DCL instead of RobustDepth | Depth estimation · DENSE-Fog light | Abs Rel: 0.316 -> 0.311 | Helped |
| use DCL instead of RobustDepth | Depth estimation · DENSE-Fog dense | Abs Rel: 0.605 -> 1.182 | Hurt |
| add BCC (ablation) | DCL depth estimation · DENSE-Fog light | Abs Rel: 0.636 -> 0.311 | Helped |
| add D_MFIR (ablation) | DCL depth estimation · DENSE-Fog light | Abs Rel: 0.320 -> 0.311 | Helped |
| add D_MDR (ablation) | DCL depth estimation · DENSE-Fog light | Abs Rel: 0.340 -> 0.311 | Helped |
| add L_pe (ablation) | DCL dehazing · GoProHazy | FADE: 0.6959 -> 0.6914 | Helped |
| add L_s (ablation) | DCL dehazing · GoProHazy | FADE: 0.8163 -> 0.6914 | Helped |
| add L_rec (ablation) | DCL dehazing · GoProHazy | NIQE: 3.7030 -> 3.4412 | Helped |
| use non-uniform β instead of constant β (ablation) | DCL · DENSE-Fog light | Abs Rel: 0.325 -> 0.311 | Helped |
Surprising findings
- On DENSE-Fog dense fog, DCL's abs Rel (1.182) is worse than several baselines (e.g., RobustDepth 0.605), and the authors attribute the inconsistency between metrics to blurred depth being closer to the mean.
Limitations
- Struggles to recover fine details of small objects such as tree branches and wires in dense haze, causing artifacts.
- Collecting large amounts of high-quality misaligned data in dynamic scenes with people and vehicles is difficult, limiting generalization.
- Dehazing-depth metrics show inconsistency in dense fog; blurred depth can score well on some metrics.
Experimental setup
Datasets: GoProHazy DrivingHazy InternetHazy DENSE-Fog REVIDE
Benchmarks: GoProHazy DrivingHazy InternetHazy DENSE-Fog REVIDE
Metrics: FADE NIQE Abs Rel RMSE log delta1 delta2 delta3 PSNR SSIM
Baselines: DCP RefineNet CDD-GAN D4 PSD RIDCP PM-Net MAP-Net NSDNet DVD MonoDepth2 MonoViT
Builds on: Monodepth2 DVD atmospheric scattering model dark channel prior
Closest papers by meaning
- Depth Information Assisted Collaborative Mutual Promotion Network for Single Image Dehazing (CVPR 2024 · similarity 0.89)
- PromptHaze: Prompting Real-world Dehazing via Depth Anything Model (AAAI 2025 · similarity 0.87)
- PRISM: Rethinking Atmospheric Scattering Reconstruction as a Unified Understanding and Restoration Model for Real-world Dehazing (NeurIPS 2026 · similarity 0.87)
- LiBrA-Net: Lie-Algebraic Bilateral Affine Fields for Real-Time 4K Video Dehazing (NeurIPS 2026 · similarity 0.87)
- PEIE: Physics Embedded Illumination Estimation for Adaptive Dehazing (AAAI 2025 · similarity 0.86)
- Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing (CVPR 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.