Vision Transformer-Based Monocular Depth Estimation for Fisheye Cameras

Monocular self-supervised depth estimation is a critical perceptual challenge in autonomous driving. With the widespread adoption of fisheye cameras in this domain, adapting self-supervised depth estimation networks to fisheye camera inputs becomes both a challenging and practical task. In this paper, we propose the MonoFVT framework, which incorporates a vision transformer (ViT) into the traditional monocular self-supervised depth estimation network, enabling it to better handle the global distortions inherent in fisheye camera images. Our novel fisheye reprojection module is versatile, compatible with various fisheye camera models, and can be easily extended to different deployment scenarios. Additionally, we introduce a scheme that utilizes speed sensor data to help the network learn absolute depth.