FutureSurf: Future Rendering ≠ Future Surface

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 points
croissant.jsonld	—	Croissant 1.0 metadata

Download

from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="rickyshi/futuresurf", filename="futuresurf_dataset.zip", repo_type="dataset", local_dir=".")

Layout (per motion, after unzipping)

dataset/controlled/<motion>/
  images/                r_0.png … r_199.png      200 renders, 800×800 RGBA
  mesh_gt/               <motion>0.ply … 199.ply  200 GT meshes (vertex-colored PLY)
  transforms_train.json  observed: 150 frames (time ≤ 0.75)
  transforms_test.json   future:    50 frames (time > 0.75)
  points3d.ply           init point cloud (Gaussian backbones)

Protocol

Train on transforms_train.json (observed 75%), then score the extracted mesh against mesh_gt at the future frames. Metric: Chamfer distance on vertex positions. See the paper and toolkit for the full protocol and scoring code.

The 8 controlled motions

Surface-changing:

• wave — periodic
• compound — periodic + trend
• stretch — monotonic
• bulge — localized
• accel — accelerating

Falsification controls:

• twist — surface-invariant
• rotate — rigid rotation
• stop — freezing

Citation

If you use FutureSurf, please cite:

@article{shi2026futuresurf,
  title   = {Future Rendering $\neq$ Future Surface: A Benchmark and Dataset for Dynamic Surface Reconstruction Beyond the Observed Window},
  author  = {Shi, Yukun and Gong, Minglun},
  journal = {arXiv preprint arXiv:2607.21471},
  year    = {2026}
}

License

CC BY 4.0