Tong He

Meta TBD Lab. - Reasoning, RL, Post-training, LLMs

RL, thinking, self-improvement, tool use, multi-modality. - Deep Think (co-creator & tech lead), AI IMO Gold Medal - Gemini 2.5 Pro (core contributor)

Perception, E2E driving, multi-modality. - Gemini-based E2E self-driving with visual CoT reasoning (core contributor) - Temporal sensor fusion for 3D object detection & segmentation (tech lead)

Advisor: Stefano Soatto # Imitation Learning: driving policy learning by latent space distillation-based Squeeze-and-Mimic networks training. # 3D Object Detection: developed the state-of-the-art 3D vehicle detection system using only a single-view image. # Point Cloud Descriptor Learning: proposed a data-driven method for joint 3D keypoint detection and descriptor learning. # SLAM: designed a grid-based uniform feature tracker to reduce the trajectory drift of visual-inertial SLAM systems.

Mentor: Pei Sun, Chenxi Liu # 3D Object Detection and Transformers (non-disclosure agreement).

Mentor: Yuanlu Xu, Shunsuke Saito, Tony Tung # Animatable 3D Object Reconstruction: photorealistic and animatable reconstruction of clothed humans in both original and canonical poses, leveraging deep implicit functions and point-bashed human shape priors.

Mentor: Hailin Jin, John Collomosse # 3D Human Pose and Shape Modeling: improved the global mesh topology regularities and local surface details of single-view clothed human reconstruction. Such enhancements are achieved by a novel implicit function formulation built upon geometry and pixel aligned features.

Mentor: Jue Wang, Haibin Huang # Point Cloud Analysis: proposed a novel deep learning method for point cloud geodesic topology estimation and utilized it to achieve the state-of-the-art performance on point upsampling, normal estimation and non-rigid shape classification.

Advisor: Shuaicheng Liu, Bing Zeng # Video Stabilization: proposed a unified optimization method for joint video stitching and stabilization via image mesh warping and bundle path video stabilization; helped to transplant it into a plug-in of Adobe After Effects. # Medical Image 3D Reconstruction: built a system for endoscopic video 3D reconstruction by Shape-from-Shading (SfS) with deformable mesh smoothness constraint.

3D object/scene representation learning, autonomous vehicle control estimation.

Yingcai Honors College (for the top 5% undergraduates in UESTC)