Abstract
Three-dimensional (3D) image reconstruction is an important field of computer vision for restoring the 3D geometry of a given scene. Due to the demand for large amounts of memory, prevalent methods of 3D reconstruction yield inaccurate results, because of which the highly accuracy reconstruction of a scene remains an outstanding challenge. This study proposes a cascaded depth residual inference network, called DRI-MVSNet, that uses a cross-view similarity-based feature map fusion module for residual inference. It involves three improvements. First, a combined module is used for processing channel-related and spatial information to capture the relevant contextual information and improve feature representation. It combines the channel attention mechanism and spatial pooling networks. Second, a cross-view similarity-based feature map fusion module is proposed that learns the similarity between pairs of pixel in each source and reference image at planes of different depths along the frustum of the reference camera. Third, a deep, multi-stage residual prediction module is designed to generate a high-precision depth map that uses a non-uniform depth sampling strategy to construct hypothetical depth planes. The results of extensive experiments show that DRI-MVSNet delivers competitive performance on the DTU and the Tanks & Temples datasets, and the accuracy and completeness of the point cloud reconstructed by it are significantly superior to those of state-of-the-art benchmarks.