Future Rendering ≠ Future Surface: A Benchmark and Dataset for Dynamic Surface Reconstruction Beyond the Observed Window

FutureSurf is a new benchmark and controlled dataset for future-time dynamic surface reconstruction, with exact future ground-truth meshes and evaluation code.

The authors show that several reconstruction methods still have large future-surface geometry errors, even when their rendered outputs look strong.

Results suggest that good novel view synthesis scores do not reliably mean accurate prediction of unseen future geometry.

The benchmark highlights a gap in current evaluation, especially for robotics and AR use cases that need trustworthy future 3D surface predictions.