Chonghyuk (Andrew) Song

I'm a 1st-year Ph.D. student at MIT's Computer Science and Artificial Intelligence Laboratory, where I am extremely fortunate to be advised by Vincent Sitzmann.

Prevously, I completed my M.S. degree at CMU's Robotics Institute, where I had the distinct pleasure of being co-advised by Deva Ramanan and Jun-Yan Zhu. I graduated from KAIST with a B.S. in Mechanical Engineering. I also interned at Snap.

My research interests lie at the intersection of machine learning and computer vision. I am especially interested in learning structure, such as 3D scenes and dynamics, from videos.

Github  /  Google Scholar  /  Twitter  /  LinkedIn  /  Email

profile photo
Total-Recon: Deformable Scene Reconstruction for Embodied View Synthesis
Chonghyuk Song, Gengshan Yang, Kangle Deng, Jun-Yan Zhu, Deva Ramanan
ICCV, 2023
code | project page | paper

Given a RGBD video containing deformable actors (top row), Total-Recon reconstructs the scene as a compositional 4D neural field, enabling extreme view synthesis from embodied views and bird's-eye views (bottom row).

Distinctiveness Oriented Positional Equilibrium for Point Cloud Registration
Taewon Min, Chonghyuk Song, Eunseok Kim, Inwook Shim
ICCV, 2021

A novel positional embedding module that yields unprecedented improvements in the accuracy of existing GNN-based rigid point cloud registration methods.

Improving Gradient Flow with Unrolled Highway Expectation Maximization
Chonghyuk Song, Eunseok Kim, Inwook Shim
AAAI, 2021

Replacing unrolled expectation maximization (EM) layers in neural networks with generalized EM layers based on the Newton-Rahpson method introduces highway connections, resulting in improved gradient flow during backpropagation.

Efficient Design Space Exploration of Multi-Mode, Two-Planetary-Gear, Power-Split Hybrid Electric Powertrains via Virtual Levers
Chonghyuk Song, Jaeho Hwang, Dongsuk Kum

A highly efficient design methodology that finds the optimal multi-mode, two-planetary-gear powertrain by leveraging the virtual lever, a modeling tool that eliminates the redundancy in the physical design space.

Credits to Jon Barron for this website's template.