Multi-scale Adaptive Region Matching Network for 3D Reconstruction

3D reconstruction can be used to satisfy the increasing requirement of people to 3D images and 3D videos with naked eye or not. With the development of deep learning, more and more attention to deep learning applications in the field of three-dimensional reconstruction of binocular vision, especially in recent years many deep learning networks are put forward to overcome most complicated calculation and low precision of the traditional three-dimensional reconstruction. Aiming at this kind of problems, put forward a kind of based on multi-scale adaptive region matching network of 3D reconstruction method, by using the adaptive feature extraction to extract the network object richness of different texture information, using multi-layer match the weight on polymerization of cost optimization, emphasize its useful information, inhibition of irrelevant information. Experimental results show that the proposed method can be used to process the binocular vision with occluded, texture-free or low-textured local backgrounds and generate higher quality 3D reconstructed objects than the previous methods. The proposed network may provided a more efficient way for 3D reconstruction.