Allen Tu

Portrait

atu1@umd.edu


IARPA Walk-through Rendering from Images of Varying Altitude (WRIVA)

Unconstrained 3D reconstruction and novel view synthesis in challenging real-world environments. PIs: Professor Tom Goldstein, Professor Matthias Zwicker, Professor Abhinav Shrivastava, Dr. Abhay Yadav, Dr. Cheng Peng, Professor Rama Chellappa

Diffusion Model Priors for 3D Reconstruction

Diffusion in the Loop produces sharper reconstructions with fewer artifacts than the baseline 3D model in sparse-view reconstruction settings. Applying posthoc diffusion cleanup further enhances visual quality and fine detail.

Publications


Boost DeformableGS rendering speed from 20 to 276 FPS using temporal sensitivity pruning and groupwise SE(3) motion distillation, all while preserving the superior image quality of per-Gaussian neural motion.




Enhance 3D Gaussian Splatting by selectively injecting super-resolution only where high-frequency detail is missing, yielding sharper results and improved perceptual quality without introducing multi-view inconsistencies.




Accelerate 3D Gaussian Splatting rendering speed by over 6× and reduce model size by over 90% through accurately localizing primitives during rasterization and pruning the scene during training, providing a significantly higher speedup than existing techniques while maintaining competitive image quality.




Prune 90% of primitives from any pretrained 3D Gaussian Splatting model using a mathematically principled sensitivity score, more than tripling rendering speed while retaining more salient foreground information and higher visual fidelity than previous techniques at a substantially higher compression ratio.


* denotes equal contribution.