Tan Guan Yuan

I am a 2nd-year PhD candidate at Monash University, focusing on 4D reconstruction and physics simulation. My goal is to build intelligent systems that perceive physical properties and causal rules from visual data, paving the way for cognitive digital twins.

I am currently exploring research internship and visiting researcher opportunities for 2026. I welcome collaborations on 3D/4D vision, neural rendering, and physics-informed learning.

Email  /  LinkedIn  /  Scholar  /  Github

Selected Publications

A spatially-factorized motion field that decouples geometry from non-rigid dynamics via a learnable, per-point kinematic subspace. Enables real-time rendering (>163 FPS), compact storage (~11 MB), and fast training (~18 min) with no SfM initialization required.

Introduces a dual-network architecture that decouples local fine-grained deformation from global motion dynamics using optical flow guidance. This approach resolves the conflict between detail preservation and temporal coherence in dynamic scene reconstruction.

A mixture-of-experts framework that orchestrates open-source LLMs to generate executable Python code for complex visual tasks. Herald achieves SOTA accuracy on various benchmarks without any model training or fine-tuning.

Proposes a self-attention deep temporal convolutional network (SADTCN) to predict non-linear tremor signals, enabling accurate, real-time active suppression for Parkinson's patients.

Teaching

• FIT5215: Deep Learning (Masters)
• FIT3181: Deep Learning

Design: Jon Barron.