This task is to estimate depth from a single color image. Estimating absolute or relative depth from camera inputs can help many applications such as scene understanding, self-driving (along with LIDAR sensors), and augmented reality (AR).
A self-supervised monocular depth estimation model. The authors propose three improvements on architecture and self-supervision loss: (1) per-pixel minimum reprojection loss, (2) auto-masking stationary pixels, and (3) multi-scale estimation.