Improving Resolution of 3D Surface With Convolutional Neural Networks

High-resolution (HR) 3D point cloud is always desired for smart city. Phase measuring profilometry (PMP) has widely used to generate 3D point cloud. However, due to the limitation of hardware, PMP is usually difficult to obtain HR 3D point cloud. This inspires us to exploit low-resolution (LR) pattern images or LR phase image to generate HR 3D point cloud. Specifically, we attempt to solve this problem using deep learning based super-resolution (SR) methods. We formulate a new deep learning based SR method for 3D point cloud. We show that the proposed SR surely improves the resolution of the reconstructed 3D point cloud. In experiments, we compare the proposed SR with other state-of-the-art SR methods, proving that the proposed SR can yield better quality of the reconstructed 3D point cloud and lower computational cost.