Dongsu Zhang | 장동수
About Me
Email: 96lives@{snu.ac.kr, gmail.com}
I’m a researcher at Seoul National University advised by Prof. Young Min Kim (currently fulfilling Korean military service duty). My research goal is to create a machine that can perceive and learn from the geometry of the spatial world that we live in and design a “better” one. I believe that this is crucial in terms of 1) it’s application, automating the design process of geometry (from everyday tools and furnitures to structures in architecture) that we leverage, and 2) building an AI system, which I think one of the greatest thing about human intelligence is the usage of tools. To do so, my main research focus has been on 3D generative models. Especially, I’ve been developing a generative model named Generative Cellular Automata (GCA), which iteratively deforms a shape locally to produce a new shape in a spatially scalable way.
I was fortunate to recieve my master’s degree advised by Prof. Young Min Kim, and bachelor’s degree in CS with math as minor in SNU. I also spent wonderful time working as a research scientist intern at NVIDIA Toronto AI Lab mentored by Amlan Kar and Sanja Fidler.
Selected Publications
For full publications list, visit my Google Scholar profile.
Investigating Chiral Morphogenesis of Gold Using Generative Cellular Automata
Sang Won Im* , Dongsu Zhang* , Jeong Hyun Han, Ryeong Myeong Kim, Changwoon Choi , Young Min Kim** , Ki Tae Nam**Nature Materials, 2024
Outdoor Scene Extrapolation with Hierarchical Generative Cellular Automata
Dongsu Zhang , Francis Williams , Zan Gojcic , Karsten Kreis , Sanja Fidler , Young Min Kim , Amlan KarCVPR, 2024 (Highlight)
pdf | code | project page
We generate a complete scene for LiDAR scans, even in regions complete unseen by LiDAR, such as above height limits of LiDAR or occluded parts.
Probabilistic Implicit Scene Completion
Dongsu Zhang , Changwoon Choi , Inbum Park , Young Min KimICLR, 2022 (Spotlight)
pdf | code
We tackle the problem of probabilistic scene completion for the first time by extending the Generative Cellular Automata to produce continuous 3D surface.
Learning to Generate 3D Shapes with
Generative Cellular Automata
Dongsu Zhang , Changwoon Choi , Jeonghwan Kim , Young Min Kim ICLR, 2021
pdf | code
We present a Markov chain based 3D generative model named Generative Cellular Automata, which is scalable for producing high-resolution voxels.