Author: Mayoore Selvarasa JaiswalPublisher:
ISBN:
Category :
Languages : en
Pages : 123
Book Description
With the introduction of economical depth cameras, computer vision research has made a huge leap forward in 3D reconstruction and understanding. However, the quality of the depth images are limited: 1) depth images contain holes and random noise due to the characteristics of the cameras and the physical world, and 2) depth images have a lower spatial resolution. These challenges make 3D object reconstruction with a limited number of RGB-D frames a difficult task. In this dissertation, we propose a method to construct 3D object models with a limited number of RGB and depth frames. We developed a complete 3D object model construction process with automatic object segmentation, pairwise registration, global alignment, model denoising, and texturing, and studied the effects of these functions on the constructed 3D object models. We also developed a process for objective performance evaluation of the constructed 3D object models. High-quality depth images are paramount to create quality 3D object models. Many depth denoising methods blur object boundaries. Partial point clouds created from depth images denoised by these methods have artifacts that make them unsuitable for 3D object modeling. We propose a method to find the clean depth edge image using the noisy depth image and a color image, then use this clean depth edge image with the proposed edge and color aware adaptive trilateral filter to obtain denoised depth images. Finally, we propose to apply depth denoising and depth super-resolution in the 3D object modeling process. We preprocess depth images with the proposed depth denoising method and a state-of-the-art depth super-resolution method, then experiment the use of these preprocessed depth images under various conditions in the 3D object modeling process. We show that depth denoising and super-resolution can improve the quality of the 3D object models.
