Deep Pruner and Adaptive Cost Volume Multiview Stereo Network for 3D Reconstruction

Reconstructing three-dimensional (3D) images is imperative in computer vision because it assists in restoring the 3D structure of a scene. However, challenges like accurate matching in low-texture and reflective areas, along with inefficient feature extraction, degrade 3D reconstruction quality and increase computational complexity. To address these challenges, we propose a robust multi-view stereo network, DPrun-RMVSNet, designed to enhance matching in occluded regions and improve feature extraction for texture-less and reflective surfaces. Our model incorporates a recurrent neural network (RNN) with long-short-term memory (LSTM) to handle depth interference. The feature network captures essential information about the image content, such as edges, textures, and corners. To reduce computational costs, we introduce a novel deep pruner feature network (DPF) with an adaptive cost volume, enabling efficient and accurate 3D model creation. The proposed model was trained using the public DTU dataset and evaluated on two benchmark datasets including DTU, and Tank and Temple. Additionally, we conduct an ablation study to assess the impact of the proposed methods, offering both quantitative and qualitative evaluations to validate the model's effectiveness. Experimental results show that our model improves state-of-the-art (SOTA) approaches, achieving better reconstruction accuracy while using less execution time and memory.