Unsupervised Fast Scene Depth Estimation Based on Laplacian Pyramid

Currently, most scene depth estimation based on deep learning relies on supervised methods, which require a large number of real labels. However, obtaining such labels in practice is often challenging. Existing unsupervised methods have problems such as low accuracy, significant artifacts, and slow inference speed. Therefore, addressing these challenges, we propose an unsupervised fast depth estimation network based on the Laplace pyramid. This network is trained by minimizing losses related to sparse matching point pair disparity using Scale-invariant Feature Transform (SIFT), image reconstruction and the consistency of left and right depth maps generated by the weight-sharing sub-network. During deployment, only one image needs to be input to complete the testing. Experimental results on the KITTI dataset demonstrate that our method outperforms the compared unsupervised depth estimation algorithms in terms of depth map accuracy, image clarity, and inference speed. For images with a resolution of $1152 \times 256$, the inference speed on an RTX2080Ti graphics card reaches 46fps.