I am a PhD candidate at the Seoul National University (SNU) Biointelligence Lab, supervised by Byoung-Tak Zhang. Previously, I graduated from the Hong Kong University of Science and Technology (HKUST) with a Bachelor's in Electronic and Computer Engineering. During my undergrad, I completed my Bachelor's thesis on Autonomous Navigation and Object Detection with Reinforcement Learning in RAM-Lab advised by Prof. Ming Liu.
I am interested in 3D vision, representation learning, and generative AI. Specifically, learning representations in a self-supervised and structured way. Most of my work is about inferring the physical world from images and videos or methods that use domain-specific knowledge of geometry and graphics. I enjoy being inspired by neuroscience to understand human intelligence of AI.
ICML, 2026
We present a kinematics-aware framework for monocular dynamic scene reconstruction that separates static and dynamic Gaussian primitives and reparameterizes their anisotropy using motion-aligned priors, enabling reconstruction from blurry videos.
To be updated
We propose a novel view-space ray grouping strategy that preserves the local geometry of Gaussians in 4D scenes via spatial constraints. By achieving physically plausible motion without external guidance, our method outperforms existing baselines in temporal consistency and reconstruction quality on monocular datasets.
IEEE Access, 2026
We propose a one-shot 3D affordance learning framework that represents affordances as 3D Gaussians by learning verb-centric affordances from a single affordance-labeled image and lifting action-aligned 2D visual features into a dense 3D representation for robotic manipulation.
ICCV, 2025
We propose OCK, a dynamic video prediction model leveraging object-centric kinematics and object slots. We introduce a novel component named Object Kinematics that comprises explicit object motions, serving as an additional attribute beyond conventional appearance features to model dynamic scenes.
CVPR NFBCC Workshop, 2025
We address the challenge of novel view synthesis from blurry inputs by generating dense 3D Gaussians, restoring sharpness, and reconstructing detailed 3D geometry of the scene affected by dynamic motion variations.
ECCV, 2024
We propose CSEConv, a novel point convolution layer equivariant under continuous SO(3) actions. Its structure is founded on the framework of group theory, realizing the convolution module defined on a sphere.
KSC, 2024 — (Best Presentation Award)
We propose a novel view synthesis method based on 3DGS that sequentially processes blurry input frames to generate new viewpoints without requiring pre-calculated camera poses. Our approach simultaneously achieves high performance in both view synthesis and camera pose estimation under dynamic conditions.
NeurIPS CRL Workshop, 2023
Object-centric representations have emerged as a promising tool for scene decomposition by providing useful abstractions. We propose a method called Object-centric Slot Patch Transformer, which is a dynamics learning mechanism that predicts future frames in an object-centric way.
CLeaR, 2023
Local independence, such as context-specific independence, has primarily been studied for discrete variables despite its importance for understanding fine-grained causal relationships. We extend this concept to continuous variables by defining its properties and introducing a differentiable method for its discovery.