Multi-view Pixel2Mesh++: 3D reconstruction via Pixel2Mesh with more images

To meet the increasing demand for high-quality 3D models, we propose an end-to-end deep learning network architecture, which can generate 3D mesh models with multiple RGB images and is different from previous methods which generate voxel or point cloud models. Unlike the single-image-based pixel2mesh network, we introduce the ConvLSTM layer to fuse perceptual features, making it possible to process multiple images simultaneously. To constrain the smoothness of 3D shapes, we design a graph pooling layer to reduce mesh structure and define a new loss function—Smooth loss. Collaborating with the graph unpooling layer in Pixel2Mesh (P2M), the graph pooling layer guarantees the mesh topology of the final 3D shapes generated. The application of Smooth loss ensures the visual appeal and structural accuracy of 3D shapes generated. Our experiments on ShapeNet dataset show that our method, compared with previous deep learning networks, can generate higher-precision 3D shapes and achieves the best on F-score and CD. In addition, due to the introduction of fusion features from multiple images, our experimental results are more convincing and credible.