Research

Image-based 3d reconstruction is one of the key research contents in 3D computer vision. However, due to the influences of occlusion, repetitive textures, etc., there usually exist holes and outliers in the reconstructed 3D point cloud for complex scenes by the existing methods in literatures. Therefore, how to exact the intrinsic relationship from a large number of 2D images and the corresponding reconstructed 3D points and then recover an incomplete 3D scene accurately, becomes an important research topic. Deep learning is to build a multi-layer neural network according to the information processing mechanism in human's brains, and it can learn an effective representation for data. This project is intended to investigate how to recover a complex 3D scene based on the theory of deep learning. The project's main contents include: (1) Computational framework on deep learning for the recovery of a complex 3D scene, and methods for designing effective network structures; (2) Methods for constructing a single-layer model in a deep network; (3) Methods for denoising the reconstructed 3D point cloud in a deep network; (4) Self-adaptive algorithm for learning the proper number of the layers and the nodes in a deep network; (5) GPU-based algorithms for training a deep network. This study is expected to provide a novel way for the recovery of a complex 3D scene.