BibTex
@inproceedings{Cai:2022:10.20380/GI2022.11,
author = {Cai, Xinhao and Coevoet, Eulalie and Jacobson, Alec and Kry, Paul},
title = {Active Learning Neural C-space Signed Distance Fields for Reduced Deformable Self-Collision},
booktitle = {Proceedings of Graphics Interface 2022},
series = {GI 2022},
year = {2022},
issn = {0713-5424},
location = {Montr{\'e}al, Quebec},
pages = {92 -- 100},
numpages = {9},
doi = {10.20380/GI2022.11},
publisher = {Canadian Information Processing Society},
}
Abstract
We present a novel method to preprocess a reduced model, training a neural network to approximate the reduced model signed distance field using active learning technique. The trained neural network is used to evaluate the self-collision state as well as the self-collision handling during real time simulation. Our offline learning approach consists of two passes of learning. The first pass learning generates positive and negative point cloud which is used in the second pass learning to learn the signed distance field of reduced subspace. Unlike common fully supervised learning approaches, we make use of semi-supervised active learning technique in generating more informative samples for training, improving the convergence speed. We also propose methods to use the learned SDF function in real time self-collision detection and assemble it in the constraint Jacobian matrix to solve the self-collision.