Calculating Depth for Dark Images Using Self-Supervised Learning: An Investigation of the Role of Edge Maps as Input

Depth estimation from a single image is a challenging problem that has received significant attention recently. In low-light conditions and with multiple light sources where the pixel intensities are not uniform, the problem becomes even more difficult due to the lack of visual information in the image. In response to the lack of suitable datasets that adequately represent these specific issues, we undertook the task of collecting our own dataset. This initiative was driven by the need for a more accurate and comprehensive representation of the problem space. Leveraging this custom dataset, which we call the dark depth dataset, we developed and applied a novel self-supervised approach to effectively address the problem. In this paper, we propose a novel method for depth estimation in low-light conditions that utilizes edge information. We have used edge information since they are prominent even in low-light conditions and gives a better understanding of the scene than simple images. Our approach is based on the self-supervised monocular depth estimation framework and extends it by incorporating edge information for improved accuracy. We have tested the edge input at different network parts to see the effect on accuracy. With this research, edge information is crucial for depth estimation as most of the points in an object are usually at the same distance, and edges can give boundaries to identify objects. However, the edge information gets lost during deep learning processes while going through different layers, so we provided the edge map at different layers to check where this edge information is most effective for the depth estimation process. Our experiment shows that giving edge information with edge-concatenation gives the best results with dark dataset. We achieved an improvement of nearly 50% over the base model in terms of $\delta $ values.