For one of the upcoming STORM seminars, Chems-Eddine Himeur will do a training for his presentation at JFIG titled : Edge detection in 3D Point-Clouds using Machine Learning.
Neural networks have gained a lot of attraction lately within the point cloud analysis community. The past year has seen an explosion in numbers in favor of Convolutional neural networks (CNN) for different analysis processes on point clouds such as segmentation, semantic segmentation and classification. Yet, simple processes like edge extraction did not witness such advancement although the subject needs more exploration. As apposed to edge extraction on images, edges on point clouds lack even a clear definition in the state of art, and their extraction also proved to be a challenging process to carry out with regular algorithms. In this paper, we formalize a new definition for edges and contours on point clouds, where both edges and their surrounding are described. We define a new parameterization scheme that turn raw point-clouds into a structured Scale-Space matrix, which contains rich features that are highly descriptive of edges. On top of this representation, we propose a novel neural network architecture that processes Scale-Space matrices to classify edges and their neighborhood. This new architecture is compact, requires small learning sets, is very fast to train and classifies millions of points in seconds. We also demonstrate that our architecture proposal called PCEDNet outperforms CNNs for edge detection in 3d point-clouds.
This seminar will take place on Thursday 07/11 at 12:30pm in room Salle des Thèses at IRIT.
(The room has been booked but has not been confirmed. The information will be updated if there is a change !)