Yu Sun

I am a Machine Learning Scientist at Meshcapade. Previously, I was a doctoral student at Harbin Institute of Technology (HIT) and an intern at JDAI CV lab in China.

I am focusing on the research of monocular 3D human motion estimation. At Meshcapade, I work with many talents in ML team, which is led by Naureen Mahmood, Talha Zaman, Michael J. Black, and Nicolas Heron.

At HIT, I am supervised by Wenpeng Gao and Yili Fu. At JDAI, I have been supervised by Qian Bao, Wu Liu, Yun Ye, Xiaolei Lv, and Tao Mei. In exploring this field, it's an honor to work with Michael J. Black.

Email  /  Google Scholar  /  Twitter  /  Github

With a holistic 5D representation, TRACE tracks the subjects presented in the first frame through time and recovers their 3D trajectories in global coordinates. It does so in one shot.

BEV adopts an imaginary Bird's-Eye-View representation to explicitly reason about depth relationships between people.

ROMP is the first one-stage method for monocular regression of multiple 3D people. It can run at over 20 FPS on a 1070Ti GPU.

DSD-SATN attempts to disentangle the pose-related features from the others (e.g. shape, camera) and learn temporal dynamics via sorting shuffled frames.

PromptHMR adopts multi-modal prompts, e.g. 2D detection, segmentation, text, etc, to promote the HPS estimation.

TokenHMR adopts a discrete-token-based representation for SMPL-based regression.

ChatPose adopts LLMs to reason about SMPL-based regression from images or textual descriptions.

WOC captures the 3D motion of users with a single camera and drives their individual 3D virtual avatars in real-time in an online chatroom.

A survey about monocular 2D and 3D human pose estimation.

The template is borrowed from this awesome website