A Multi-Task Learning Approach for Stereo Depth Estimation

Due to the uneven depth scale in the linear disparity space, calculating stereo disparity and subsequently converting it into depth, although widely used, leads to nonlinear error amplification. Moreover, the limited availability of depth annotations in stereo datasets has impeded the progress of end-to-end depth estimation techniques. This paper introduces a semi-supervised method for depth estimation using a multitask network. The multi-task network comprises two branches for disparity estimation and depth estimation. During training, it leverages a pre-trained stereo disparity network to provide dense depth self-supervision, expediting the training of the depth branch. This network offers an effective solution to the scarcity of stereo depth datasets and the sparsity of depth information in point cloud annotations. The efficacy of the algorithm is validated on the KITTI 3D object dataset using sparse point clouds as depth annotations, showcasing remarkable depth estimation capabilities. Additionally, the paper transforms obtained depth maps into pseudo-lidar for 3D object detection, achieving promising results on the KITTI dataset.