Xu Cao 曹旭

Hi there! I am a research scientist in the graphics team at CyberAgent AI Lab. Previously, I worked as a Specially Appointed Assistant Professor (特任助教) in Computer Vision Lab. at Osaka University from 2022 to 2023.

I earned my Ph.D. (Apr. 2019 - Mar. 2022) from Osaka University under the supervision of Prof. Yasuyuki Matsushita. Prior to that, I received M.S. degree from Nagoya University in Mar. 2019 and B.S. degree from Nanjing University of Aeronautics and Astronautics (NUAA) in Jun. 2016.

My research interest includes 3D and physics-based vision.

Email  |  Google Scholar  |  Github  |  LinkedIn  |  Twitter

"It can scarcely be denied that the supreme goal of all theory is to make the irreducible basic elements as simple and as few as possible without having to surrender the adequate representation of a single datum of experience." (Albert Einstein, 1933)

Highlighted are my favorite works.

SuperNormal efficiently recovers high-fidelity shapes using multi-view normal maps estimated by photometric stereo. We accelerate neural SDF training by directional finite difference (DFD) and patch-based sampling.

We find a strictly held consistency in multi-view stereo when using azimuth maps as inputs and leverage this consistency for neural SDF optimization.

ECON is designed for full-body human digitization from a single-view color image and features robust performance on in-the-wild images with loose clothing or challenging poses.

BiNI is a variational approach for recovering the surface from a single-view normal map, which is capable of preserving depth discontinuities.

Using a stereo camera and a flashlight that are common on modern mobile phones can recover high-fidelity surfaces.

Plane fitting is useful for normal estimation; plane fitting is now useful for normal integration.

Flashed stereo camera, recovered fine shapes.

Learnt features from PointNet can stylize point clouds.

Website adapted from Jon Barron.