Self-Supervised Monocular Depth Learning in Low-Texture Areas

For the task of monocular depth estimation, self-supervised learning supervises training by calculating pixel difference between target image and warped reference image, obtaining results comparable to those with full supervision. However, problematic pixels in low-texture regions are ignored, since most researchers think that no violate assumption camera motion, taking stereo pairs as input learning, which leads optimization problem these regions. To tackle this problem, we perform photometric loss using lowest-level feature maps instead implement first- second-order smoothing depth, ensuring consistent gradients ring optimization. Given shortcomings ResNet backbone, propose a new estimation network architecture improve edge location accuracy obtain clear outline information even smoothed boundaries. acquire more stable reliable quantitative evaluation results, introce virtual data set because have dense corresponding pixel. We achieve performance exceeds prior methods on both Eigen Splits KITTI VKITTI2 sets input.