Abstract
Self-supervised monocular depth estimation refers to training a monocular depth estimation (MDE) network using only RGB images to overcome the difficulty of collecting dense ground truth depth. Many previous works addressed this problem using depth classification or depth regression. However, depth classification tends to fall into local minima due to the bilinear interpolation search on the target view. Depth classification overcomes this problem using pre-divided depth bins, but those depth candidates lead to discontinuities in the final depth result, and using the same probability for weighted summation of color and depth is ambiguous. To overcome these limitations, we use some predefined planes that are parallel to the ground, allowing us to automatically segment the ground and predict continuous depth for it. We further model depth as a mixture Laplace distribution, which provides a more certain objective for optimization. Previous works have shown that MDE networks only use the vertical image position of objects to estimate the depth and ignore relative sizes. We address this problem for the first time in both stereo and monocular training using resize cropping data augmentation. Based on our analysis of resize cropping, we combine it with our plane definition and improve our training strategy so that the network could learn the relationship between depth and both the vertical image position and relative size of objects. We further combine the self-distillation stage with post-processing to provide more accurate supervision and save extra time in post-processing. We conduct extensive experiments to demonstrate the effectiveness of our analysis and improvements.
Abstract (translated)
URL
https://arxiv.org/abs/2210.01612
