Runfeng Li

I am a M.Sc. student in Computer Science at Brown University. I am very fortunate to be advised by Prof. James Tompkin.

Previously, I got my B.S. in Maths & CS from University of Illinois Urbana-Champaign, where I got some early exposure to research in human motion modeling and deep learning, advised by Prof. Liangyan Gui.

Email  /  CV  /  Github  /  Scholar

Research

I have primarily focused on 3D/4D reconstruction, studying how to physically model raw continuous-wave time-of-flight (C-ToF) and RGB signals to recover physical properties like geometry, appearance, motion, and elasticity.

We adapt Gaussian splatting for dynamic continuous-wave time-of-flight radiance fields reconstruction and propose two simple optimization heuristics that address the discrepancy problem between the rendered mean depth and the depth from rendered raw ToF frames, enabling 100x faster training, >100Hz rendering speed, and more accurate depth.

We test various monocular dynamic Gaussian splatting methods on various datasets and found that 3D Gaussian optimization is always brittle but deformation MLP based methods (smooth motion) helps.

Projects

We reimplemented PhysGaussian using Taichi, and explored Gaussian elasticity (Young's modulus) estimation through the differentiable MPM using multi-view videos.

I adapted Gaussian Splatting for HDR reconstruction of RawNeRF scenes, achieved good results on rendered raw images but encountered increased reconstruction errors after tonemapping, which might be due to insufficient training or need more careful hyperparameter tuning.

I implemented single and multi-grid Poisson PDE solvers for the gradient domain HDR compression method in CUDA/C++, achieving real-time (100-200Hz) HDR tonemapping for 1k-2k resolution images.