|M.Sc Student||Urbach Dahlia|
|Subject||DPDist: Comparing Point Clouds using Deep Point Cloud|
|Department||Department of Electrical and Computers Engineering||Supervisor||PROF. Michael Lindenbaum|
|Full Thesis text|
We introduce a new deep learning method for point cloud comparison. Our approach, named Deep Point Cloud Distance (DPDist), measures the distance between the points in one cloud and the estimated surface from which the other point cloud is sampled. The surface is estimated locally using the 3D modified Fisher vector representation and an implicit neural function. The 3D modified Fisher vector is based on a Gaussians’ grid, and therefore, enables a straightforward extraction of a local subgrid representation vector. The implicit function is then trained to outputs an estimated local continuous distance function of the surface regarding the input local representation vector. The local representation reduces the complexity of the surface, enabling effective learning, which generalizes well between object categories. We test the proposed distance in challenging tasks, such as similar object comparison and registration. We show that our method provides significant improvements over commonly used distances such as Chamfer distance, Earth mover’s distance, and others.
Additionally, we provide an analysis of the Chamfer distance behavior for different sampling regimes. The Chamfer distance is based on point-to-point distances, and therefore, it is common to assume that its performance strongly depends on the sampling density. However, we show that, in the context of registration tasks, the sampling regime is even more critical than the density.