Tian Zheng



Tsinghua University, Beijing, China

Email : zhengt19@mails.tsinghua.edu.cn / Phone : +86 15651722980

I am currently a graduate student at Tsinghua University. My research interest is 3D vision. I have experience in dense 3D reconstruction, 3D semantic segmentation and 3D instance segmentation. I have strong backgrounds on C++ / Python programming and computer vision / deep learning knowledge.


2019 – 2022

M.Phil., Data Science & Information Technology, TSINGHUA UNIVERSITY

Researcher at the Smart Imaging Lab of Tsinghua-Berkeley Shenzhen Institute(TBSI)

Mentor: Lu Fang

2015 – 2019

B.Eng., Electronic Engineering, SOUTHEAST UNIVERSITY

GPA: 3.96/4

Award & Honor: 2017 National Scholarship, 2018 Merit Student of Jiangsu Province


DEC. 2019 – AUG. 2020

BuildingFusion: Semantic-aware Structural Building-scale 3D Reconstruction

(TPAMI 2020)

Tian Zheng*, Guoqing Zhang*, Lei Han*, Lan Xu and Lu Fang; (*equal contribution)

In this work, we built a scalable RGBD dense reconstruction system that is able to collaboratively reconstruct building-level scenes and provide online structure and semantic understanding, achieving the current state-of-the-art. My job includes:

- Designed the room-level loop closure detection method, which is the key to scalable and robust reconstruction. Achieved 2x better recall than the previous STOA.

- Integration of real-time 3D semantic / instance networks.

- Participated in the system pipeline design.

MAR. 2019 – NOV. 2019

OccuSeg: Occupancy-aware 3D Instance Segmentation (CVPR 2020)

Lei Han, Tian Zheng, Lan Xu, Lu Fang

In this work, we introduced a novel occupancy signal in the 3D instance segmentation pipeline, followed by a carefully designed clustering method. Our method ranks the 1st in the ScanNet 3D semantic and instance segmentation benchmark1,2. My job includes:

- Participated in the design of the network. Explored the best combination of network tasks.

- Implemented the network and the proposed clustering method using PyTorch C++ APIs

- Integrated the pipeline with the reconstruction system, resulting in a live 3D reconstruction / perception system.