Triaxial Squeeze Attention Module and Mutual-Exclusion Loss Based Unsupervised Monocular Depth Estimation

Monocular depth estimation plays a crucial role in scene perception and 3D reconstruction. Supervised learning based depth estimation needs vast amounts of ground-truth depth data for training, which seriously restricts its generalization. In recent years, the unsupervised learning methods without LiDAR points cloud have attracted more and more attention. In this paper, an unsupervised monocular depth estimation method using stereo pairs for training is designed. We present a triaxial squeeze attention module and introduce it into our unsupervised framework to augment the representations of the depth map in detail. We also propose a novel training loss that enforces mutual-exclusion in image reconstruction to improve the performance and robustness in unsupervised learning. Experimental results on KITTI show that our method not only outperforms existing unsupervised methods but also achieves better results comparable with several supervised approaches trained with ground-truth data. The improvements in our method can better preserve the details of the depth map and allow the shape of objects to be maintained more smoothly.