Depth-Centric Dehazing and Depth-Estimation from Real-World Hazy Driving Video

Authors: Junkai Fan, Kun Wang, Zhiqiang Yan, Xiang Chen, Shangbing Gao, Jun Li, Jian Yang

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate the proposed method outperforms current state-of-the-art techniques in both video dehazing and depth estimation tasks, especially in real-world hazy scenes. We compare our method against state-of-the-art image/video dehazing and depth estimation techniques. Additionally, we perform ablation studies to highlight the impact of our core modules and loss functions.
Researcher Affiliation Academia 1PCA Lab, Key Lab of Intelligent Perception and Systems for High-Dimensional Information of Ministry of Education, School of Computer Science and Engineering, Nanjing University of Science and Technology, China 2Huaiyin Institute of Technology, China EMAIL, EMAIL
Pseudocode No The paper describes the methodology using textual explanations and mathematical equations, but it does not include a clearly labeled pseudocode or algorithm block.
Open Source Code Yes Code https://github.com/fanjunkai1/DCL
Open Datasets Yes We evaluate the proposed method separately using video dehazing datasets (e.g., Go Pro Hazy, Driving Hazy and Internet Hazy) (Fan et al. 2024) and depth estimation datasets (e.g., DENSE-Fog) (Bijelic et al. 2020) in real hazy scenes.
Dataset Splits Yes Go Pro Hazy, consists of videos recorded with a Go Pro 11 camera under hazy and clear conditions, comprising 22 training videos (3791 frames) and 5 testing videos (465 frames). In contrast, Driving Hazy was collected... This dataset contains 20 testing videos (1807 frames)... Specifically, we used 572 dense-fog images and 633 light-fog images to assess all depth estimation models.
Hardware Specification Yes Our model is trained for 50 epochs using Py Torch on a single NVIDIA RTX 4090 GPU, with training taking approximately 15 hours on the Go Pro Hazy dataset. We compared the parameter count, FLOPs, and inference time of the SOTA methods for image/video dehazing and self-supervised depth estimation tasks on an NVIDIA RTX 4090 GPU.
Software Dependencies No The paper mentions using Py Torch and the ADAM optimizer but does not specify their version numbers.
Experiment Setup Yes The initial learning rate is set to 1e 4 and decays by a factor of 0.1 every 15 epochs. The batch size is 2, and the input frame size is 640 192. Our model is trained for 50 epochs... The final loss parameters are set as follows: η = 1e 1, γ = 2e 1, ξ = 1e 3, ω1 = 4e 3 and ω2 = 1e 3.