Relightables: A general 3D reconstruction pipeline in Google

Release Time:2020-06-10Number of visits:359

Speaker:   Dr. Kaiwen Guo

Time:       10:00-11:30, Jun. 12

Location:  SIST 1A 200

Host:       Prof. Lan Xu

 

Abstract:

In this talk, I will introduce a general photogrammetry system built in Google. Relightables is based on the latest generation of LightStage with more than 100 high resolution 4K infrared and RGB cameras used to capture synchronized human performance. The system generates raw data at 60 Hz with throughput of 65.3 GB/s cached at a local datacenter. We customized a novel distributed software pipeline using Google Cloud to process data in parallel, generating extremely high-quality results automatically for both static characters modelling and performance capture, better than any photogrammetry software in the market. Besides, I will also talk about depth-fusion based human modeling based on a series of works since my PhD and share a few thoughts on geometric capture with deep neural networks.

Bio:

Kaiwen Guo is a research scientist in Google focusing on 3D computer vision and graphics. He is an active researcher and published papers on SIGGRAPH ASIA, CVPR, ICCV, ECCV, 3DV, TOG, TPAMI and TVCG. He acquired PhD on computer science from Tsinghua University, supervised by Prof. Qionghai Dai and Prof. Yebin Liu,  and BS on automation from Northeastern University.

SIST-Seminar 18240