Distortion-Aware Depth Self-Updating for Self-Supervised Fisheye Monocular Depth Estimation.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Fisheye cameras offer a wide field of view, crucial for applications like autonomous driving.
• Existing self-supervised monocular depth estimation methods struggle with fisheye image distortions.
• Accurate depth perception is vital for scene understanding and navigation.

Purpose of the Study:

• To develop a robust self-supervised monocular depth estimation method for fisheye cameras.
• To overcome the limitations posed by severe image distortions in fisheye lenses.
• To enhance depth estimation accuracy in challenging wide-angle imaging scenarios.

Main Methods:

• Proposed DDS-Net (distortion-aware depth self-updating network) utilizing a coarse-to-fine learning strategy.
• Introduced a distortion-aware fisheye cost volume construction module for accurate feature matching.
• Implemented a depth self-updating module for iterative refinement of depth maps.

Main Results:

• DDS-Net significantly outperforms 14 state-of-the-art methods on three fisheye datasets.
• The distortion-aware cost volume effectively captures pixel-level depth cues despite severe distortions.
• Iterative depth map updating enhances estimation accuracy.

Conclusions:

• DDS-Net provides a significant advancement in self-supervised fisheye monocular depth estimation.
• The proposed modules effectively mitigate distortion issues inherent in fisheye imagery.
• This method offers a more reliable solution for depth perception with wide-angle cameras.