Indoor Monocular Image Depth Estimation Based on Semantic Information of Tree-shaped ASPP Structure

Abstract

ASPP (Atrous Spatial Pooling Pyramid) has the advantage that it can expand the receptive field and extract multiscale features without changing the image resolution. We introduce it into the depth estimation task to improve the problems of inaccurate depth estimation, blurred edges, and loss of depth information details on the current unsupervised depth estimation methods for indoor monocular images. However, the ASPP module does not consider the relationship between different pixel features, resulting in inaccurate extraction of scene features in the depth estimation task. Therefore, we propose a Tree-shaped ASPP structure for this drawback, combined with the SC-SfMLearner network using the NYUv2 dataset, adding the spatial semantic information pool formed by the ASPP tree structure between the encoder and decoder structures of the depth estimation network, which can not only expand the receptive field without losing resolution but also capture and fuse multi-scale context information, so that different pixels establish connections between features. The results show that, compared with the original method, the improved method has stronger network feature extraction ability, clearer contours of each target in the scene, more distinct layers, and more accurate depth estimation results.