标题(TITLE): Octree-based Sparse Convolutional Neural Networks for 3D Shape Analysis and Synthesis
讲者(SPEAKER): Peng-Shuai Wang
主持(HOST): 人工智能学院，姜海勇
时间(TIME): 12月14日 (周二), 19:00
地点(VENEU): 腾讯会议号790-824-553
报告摘要（ABSTRACT）：
Analyzing and synthesizing 3D shapes are of great importance for many graphics and vision applications, such as robotics, autonomous driving, virtual reality, and mixed reality. In this talk, I will present my efforts on data-driven 3D shape analysis and synthesis with deep learning, specifically Octree-based sparse convolutional networks (O-CNN). In O-CNN, we only apply the convolution to the voxels containing non-zero shape surface signals, instead of applying the convolution operation to the whole 3D space, and we use Octree to efficiently organize sparse voxels. In this way, we improve the speed and memory cost of 3D CNNs by 10 to 100 times compared with previous volumetric CNNs and achieve top performances on a series of shape analysis tasks and shape generation tasks. And to reduce the cost of label annotation for 3D CNNs, we propose a generic backbone network and a novel unsupervised 3D pretraining method. After pretraining on a large unlabeled dataset, we can finetune the network on both shape-level and point-level analysis tasks and achieve much better performance than supervised learning with limited annotations.
讲者简介（BIO）：
Peng-Shuai Wang is currently a senior researcher at MSRA. He received his Bachelor’s degree and Ph.D. degree from Tsinghua University in 2013 and 2018. His research interest includes computer graphics and 3D deep learning. He has published 7 papers as the first author on top conferences, including SIGGRAPH (ASIA), CVPR, AAAI, and IJCAI. According to google scholar, his paper on Octree-based sparse CNNs is one of the Top-5 Most Cited Papers within the past 5 years among all papers published in SIGGRAPH(ASIA) and ACM Transactions on Graphics. He was awarded the Outstanding Ph.D. Graduate of Beijing in 2018. He servers as paper reviewers for SIGGRAPH(ASIA), TVCG, TPAMI, etc., and conference program committee members of 3DV 2020 and SMI 2021.
