ParSeNet: A Parametric Surface Fitting Network for 3D Point Clouds

Abstract

We propose a novel, end-to-end trainable, deep network called ParSeNet that decomposes a 3D point cloud into parametric surface patches, including B-spline patches as well as basic geometric primitives.ParSeNet is trained on a large-scale dataset of man-made 3D shapes and captures high-level semantic priors for shape decomposition. It handles a much richer class of primitives than prior work, and allows us to represent surfaces with higher fidelity. It also produces repeatable and robust parametrizations of a surface compared to purely geometric approaches. We present extensive experiments to validate our approach against analytical and learning-based alternatives. ParseNet EditFig. 1: ParSeNet decomposes point clouds (top row) into collections of seamlessly assembled parametric surface patches including B-spline patches (bottom row). On the right, a shape is edited using the inferred parametrization.

Spatial Calibration of Humanoid Robot Flexible Tactile Skin for Human–Robot Interaction

“…The generation of the B-spline surface has been tested on the mesh of the right wrist of a HRP-4C humanoid robot. It shows approximation errors similar to what can be obtained with common NLLS approximation [24] or deep-learning-based methods [29][30][31]. It has been observed that the error is higher where the mesh has higher curvature.…”

“…The proposed method requires a user-defined set of planes and a 3D mesh as input. Thus, it is not as automatic as other surface approximation methods [24,[29][30][31]; however, it allows the user to finely tune the calibration, even for complex shapes. Currently, the method has been validated on a smooth surface but can take advantage of its configuration capabilities to be used on more complex surfaces.…”

“…For example, the approach of X. Liu et al is adapted to surfaces of revolution [27], while the one of S. Wang et al targets high-speed train heads [28]. More recently, some methods have been developed to fit a mesh of any shape to a parametric surface, such as deep neural networks that learned the surface-fitting operation from a large dataset of meshes [29][30][31]. These methods-especially [29]-show good results for meshes of any shape, but the obtained surface is hard to adjust to the needs of the user.…”

“…More recently, some methods have been developed to fit a mesh of any shape to a parametric surface, such as deep neural networks that learned the surface-fitting operation from a large dataset of meshes [29][30][31]. These methods-especially [29]-show good results for meshes of any shape, but the obtained surface is hard to adjust to the needs of the user. These limitations motivated the development of a new method for spatial calibration.…”

Primitive Shape Recognition for Object Grasping

“…More recent progress further demonstrating the potential of primitive shapes introduces different deep neural network designs such as a vowel-based network to generate primitives (Tulsiani et al 2017), an image-based network combined with Conditional Random Field (CRF) (Kalogerakis et al 2017), and point-based network with differentiable primitive model estimator ). However, those methods generally took a 3D model or a point cloud randomly sampled on a model's surface as the input (Paschalidou et al 2019(Paschalidou et al , 2020, even required fine-grained labels to train the networks (Sharma et al 2020), which is not as feasible to achieve for grasping.…”