Three-dimensional Reconstruction has drawn much attention in computer vision. Generating a dense point cloud from a single image is a more challenging task. However, generating dense point clouds directly costs expensively in calculation and memory and may cause the network hard to train. In this work, we propose a two-stage training dense point cloud generation network. We first train our attention-based sparse point cloud generation network to generate a sparse point cloud from a single image. Then we train our dense point cloud generation network to densify the generated sparse point cloud. After combining the two stages and finetuning, we obtain an end-to-end network that generates a dense point cloud from a single image. Through evaluation of both synthetic and real-world datasets, we demonstrate that our approach outperforms state of the art works in dense point cloud generation.

The research field of reconstructing 3D models from 2D images is becoming more and more important. Existing methods typically perform single-view reconstruction or multi-view reconstruction utilizing the properties of recurrent neural networks. Due to the self-occlusion of the model and the special nature of the recurrent neural network, these methods have some problems. We propose a novel threedimensional fusion hierarchical reconstruction method that utilizes a multi-view feature combination method and a hierarchical prediction strategy to unify the single view and any number of multiple views 3D reconstructions. Experiments show that our method can effectively combine features between different views and obtain better reconstruction results than the baseline, especially in the thin parts of the object.