See the Unseen from Single-view Videos
A unified end-to-end differentiable framework that simultaneously recovers force fields and material fields from a single video. We combine a 3D Gaussian Splatting reconstruction of the first frame with a differentiable Material Point Method simulator, with material parameters initialized by a Vision-Language Model. The whole pipeline is optimized directly from the input video using only pixel-level reconstruction losses (MSE + SSIM) — no proxy supervision required.
By Jun Dai and Sheng Zhao — Machine Vision Course Project, University of Rochester.
Scan the QR code, or click here to open the project page.
This repository hosts the source of the project page deployed via GitHub Pages. It contains:
index.html— the project pagestatic/css/— page stylesstatic/images/— teaser, pipeline figure, and result figuresstatic/pdfs/— the paper PDFstatic/ppt/— presentation slide decks
The source code of the method itself is maintained separately.