|A Data Drive Approach for 3D Segmentation|
|Date: 2017/4/18 Browse: 64|
Speaker: Ye Duan
In this talk, I will talk about two 3D segmentation algorithms that we recently developed. I will first talk about a primitive-based 3D segmentation algorithm for mechanical CAD models (represented by either mesh or point cloud). Our proposed approach differs from existing techniques in the following aspects. First, by assuming that common mechanical models only have a limited number of dominant orientations that their primitives are either parallel or orthogonal to, we narrow down the search space for detecting the primitives to the automatically estimated major orientations of the input model. Second, we employ a dimension reduction method that transforms the problem of detecting 3D primitives into the classical 2D problems such as circle and line detection in images. Third, we generate an over-complete set of primitives and formulate the segmentation as a set cover optimization problem. We demonstrate our method’s robustness to noise and show that it compares favorably with state-of-the-art solutions on many synthetic and real scanned examples. In the second part of the talk I will introduce our latest work of machine learning based 3D mesh segmentation using multi-view convolutional neural networks (CNN). Evaluations on the Princeton Segmentation Benchmark dataset show that our framework significantly outperforms other state-of-the-art CNN methods.
Ye Duan (段晔) is an Associate Professor of Computer Science at University of Missouri. His research focuses on Computer Graphics, Computer Vision and Biomedical Imaging. He received his BS Degree of Mathematics from Peking University, and MS and PhD Degrees of Computer Science from State University of New York at Stony Brook.