Self-supervised monocular depth estimation with learned-depth-prior-based disentanglement of moving objects from camera ego motion

Abstract

In this paper, we address the moving object problem in self-supervised learning of monocular depth estimation. A neural network learns to estimate depth maps based on the rigid scene assumption in self-supervised monocular depth estimation. However, moving objects that violate the rigid scene assumption induce incorrect depth estimation. We point out this problem comes from the pose network that only estimates the camera ego-motion between sequential frames. Thus, we suggest a Moving Object Disentangling Network, dubbed MODNet, that estimates the camera ego-motion as well as the object motion. To induce clear boundaries of the objects in motion estimation, we add an auxiliary branch to predict the binary foreground segmentation maps on the decoder of the motion estimation network. We show our MODNet resolves the moving object problem and improves the depth estimation performance with extensive experiments on the KITTI dataset.