Research map / Video / Video segmentation and tracking
Driving-Video Dehazing with Non-Aligned Regularization for Safety Assistance
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 a non-aligned regularization approach for driving-video dehazing that matches clear reference frames from misaligned clear videos to supervise a flow-guided, cosine-attention video dehazing network, and introduces GoProHazy/DrivingHazy datasets.
Problem
Real driving-video dehazing lacks precisely aligned hazy/clear video pairs for training, because speed, path, and moving objects cause temporal and spatial misalignment.
Approach
Match each hazy frame to the most similar clear frame from a non-aligned clear video using an adaptive sliding window over VGG feature cosine similarity, then supervise the dehazing network with a multi-frame contextual reference loss; the video network aligns frames with a flow-guided cosine attention sampler and fuses them with deformable cosine attention.
Key results
- On GoProHazy, DVD achieves FADE 0.7598 and NIQE 3.7753, beating the best prior method (RIDCP FADE 0.8010; PM-Net NIQE 4.1211).
- On DrivingHazy, DVD achieves FADE 0.8745 and NIQE 3.7480, improving on PM-Net by FADE 0.1227 and NIQE 0.3119 (per the text).
- Replacing NRFM matching with unpaired random matching raises FADE from 0.7598 to 0.9204 on GoProHazy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DVD instead of NSDNet | video dehazing · GoProHazy | FADE: 0.7996 -> 0.7598 | Helped |
| use DVD instead of PM-Net | video dehazing · GoProHazy | NIQE: 4.1211 -> 3.7753 | Helped |
| use DVD instead of NSDNet | video dehazing · DrivingHazy (NoRef) | FADE: 0.9348 -> 0.8207 | Helped |
| use DVD instead of PM-Net | video dehazing · DrivingHazy (NoRef) | NIQE: 3.8944 -> 3.5825 | Helped |
| use DVD instead of NSDNet | video dehazing · InternetHazy | FADE: 0.8934 -> 0.8745 | Helped |
| use DVD instead of NSDNet | video dehazing · InternetHazy | NIQE: 4.3835 -> 3.7480 | Helped |
| add NRFM instead of randomly matching clear reference frames (ablation) | DVD · GoProHazy | FADE: 0.9204 -> 0.7598 | Helped |
| add NRFM instead of randomly matching clear reference frames (ablation) | DVD · GoProHazy | NIQE: 3.9729 -> 3.7753 | Helped |
| add FCAS to the baseline alignment module (ablation) | DVD · DrivingHazy | FADE: 0.8957 -> 0.8869 | Helped |
| add FCAS to the baseline alignment module (ablation) | DVD · DrivingHazy | NIQE: 3.9217 -> 3.8495 | Helped |
| replace non-local fusion with DCAF (ablation) | DVD · DrivingHazy | FADE: 0.8957 -> 0.8464 | Helped |
| replace non-local fusion with DCAF (ablation) | DVD · DrivingHazy | NIQE: 3.9217 -> 3.6981 | Helped |
Surprising findings
- Video-based supervised methods PM-Net and MAP-Net, retrained with contextual loss on the non-aligned data, were outperformed by the image-based non-aligned NSDNet on several metrics, suggesting alignment-based supervised video models do not transfer well to misaligned data.
Limitations
- Evaluated mainly on real-world datasets without ground truth, relying on no-reference metrics (FADE, NIQE) and user votes.
- Sampling kernel size must be tuned to motion magnitude; large kernels were limited by computational constraints.
- Datasets are small (22 training videos) and mostly urban roads with limited high-speed data.
Experimental setup
Datasets: GoProHazy REVIDE
Benchmarks: GoProHazy DrivingHazy InternetHazy
Metrics: FADE NIQE user study votes
Baselines: DCP RefineNet CDD-GAN D4 PSD RIDCP PM-Net MAP-Net NSDNet
Builds on: NSDNet SpyNet contextual loss deformable convolution VGG-16
Closest papers by meaning
- Depth-Centric Dehazing and Depth-Estimation from Real-World Hazy Driving Video (AAAI 2025 · similarity 0.84)
- LiBrA-Net: Lie-Algebraic Bilateral Affine Fields for Real-Time 4K Video Dehazing (NeurIPS 2026 · similarity 0.83)
- Frequency-aware Event-based Video Deblurring for Real-World Motion Blur (CVPR 2024 · similarity 0.82)
- Adaptive Window Pruning for Efficient Local Motion Deblurring (ICLR 2024 · similarity 0.82)
- DC-Gaussian: Improving 3D Gaussian Splatting for Reflective Dash Cam Videos (NeurIPS 2024 · similarity 0.81)
- CoA: Towards Real Image Dehazing via Compression-and-Adaptation (CVPR 2025 · similarity 0.81)
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.