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

Abstract: In this paper, we study the challenging problem of simultaneously removing haze and estimating depth from real monocular hazy videos. These tasks are inherently complementary: enhanced depth estimation improves dehazing via the atmospheric scattering model (ASM), while superior dehazing contributes to more accurate depth estimation through the brightness consistency constraint (BCC). To tackle these intertwined tasks, we propose a novel depth-centric learning framework that integrates the ASM model with the BCC constraint. Our key idea is that both ASM and BCC rely on a shared depth estimation network. This network simultaneously exploits adjacent dehazed frames to enhance depth estimation via BCC and uses the refined depth cues to more effectively remove haze through ASM. Additionally, we leverage a non-aligned clear video and its estimated depth to independently regularize the dehazing and depth estimation networks. This is achieved by designing two discriminator networks: $D_{MFIR}$ enhances high-frequency details in dehazed videos, and $D_{MDR}$ reduces the occurrence of black holes in low-texture regions. Extensive experiments demonstrate that the proposed method outperforms current state-of-the-art techniques in both video dehazing and depth estimation tasks, especially in real-world hazy scenes. Project page: https://fanjunkai1.github.io/projectpage/DCL/index.html.

• The paper introduces a depth-centric framework that jointly optimizes video dehazing and depth estimation by integrating the Atmospheric Scattering Model and Brightness Consistency Constraint.
• It leverages a shared depth network and dual discriminators to refine high-frequency details in dehazed frames and correct depth inaccuracies.
• Experiments demonstrate state-of-the-art performance on datasets like GoProHazy and DrivingHazy, enhancing visual perception for autonomous driving systems.

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

This paper introduces an advanced framework designed to simultaneously tackle video dehazing and depth estimation in real-world hazy driving videos. The authors address these interconnected tasks using a depth-centric learning approach that integrates the Atmospheric Scattering Model (ASM) and the Brightness Consistency Constraint (BCC). This integration enables the model to exploit the inherent dependencies between dehazing and depth estimation to achieve superior results in challenging real-world conditions.

Key components of the proposed approach include a shared depth estimation network leveraged by both ASM and BCC. This network processes dehazed neighboring frames to improve depth estimation via the BCC, which in turn enhances haze removal through the ASM. A significant innovation in this work is the use of two discriminator networks: $D_\text{MFIR}$ and $D_\text{MDR}$. $D_\text{MFIR}$ refines high-frequency details in dehazed frames, while $D_\text{MDR}$ addresses inaccuracies in depth estimations, often manifesting as black holes in regions lacking texture.

The extensive experiments presented demonstrate the substantial gains achieved by this approach. On various real-world datasets, the proposed Depth-Centric Learning (DCL) framework generates state-of-the-art results, clearly surpassing existing methods in both video dehazing and depth estimation. Notably, improvements are highlighted across several key datasets including GoProHazy, DrivingHazy, InternetHazy, and DENSE-Fog, where DCL consistently outperforms traditional methods.

The implications of this research are twofold. Practically, it enhances visual applications critical for autonomous systems, such as object detection and semantic segmentation under adverse weather conditions. Theoretically, the integration of ASM and BCC within a single learning framework offers a robust strategy to disentangle and solve tasks that are traditionally considered individually. This places the work in a pivotal position for the development of more comprehensive environmental perception models, particularly in automotive AI. Future directions could explore the extension of this framework to additional weather conditions or dynamic scenarios, potentially utilizing self-supervised or semi-supervised strategies to further eliminate the need for curated datasets.

Overall, by effectively bridging the tasks of dehazing and depth estimation, this research enriches the toolkit available for navigating and understanding fog-laden environments, thus contributing to safer and more reliable autonomous systems.