A benchmark for future-time surface reconstruction: train on monocular frames up to time T (the observed 75%), then score the reconstructed surface mesh at held-out future times t > T (the remaining 25%).

📄 Paper: arXiv:2607.21471 💻 Code / toolkit: github.com/Ricky-S/futuresurf

	Files

file size contents

futuresurf_dataset.zip ~0.9 GB 8 controlled motions: renders + per-frame GT meshes + transforms + init… See the full description on the dataset page: https://huggingface.co/datasets/rickyshi/futuresurf.