asergiu/HumanStereoPreview

HumanStereo Preview (< 4 GB sample)

A small sample of HumanStereo — HumanStereo: A Benchmark for Metric Facial Depth Estimation and Its Evaluation (Alexandru I. Marinescu, Diana L. Borza, Adrian S. Dărăbant; NeurIPS 2026, Evaluations & Datasets Track).

This repository exists so that the data can be inspected and tried without downloading the full 1.2 TB release. It contains a subset of rendered images of the training-split avatars (the label column in the viewer is the avatar name). For the complete dataset — 417,280 calibrated stereo sessions with metric depth, surface normals, masks, stereo calibration and the split manifest — use the main repository:

• 🤗 Full dataset: https://huggingface.co/datasets/asergiu/HumanStereo
• 💻 Code (loader, metrics, benchmark): https://github.com/asergiu-ubb/HumanStereoNeurIPS2026Public

Everything stated in the main dataset card (file formats, splits, demographics, intended and prohibited uses, ethics, maintenance) applies to this preview as well. The data are synthetic renders of commercially licensed RenderPeople scans and contain no real personal data. License: CC BY-NC 4.0.

Citation

@inproceedings{marinescu2026humanstereo,
  title     = {HumanStereo: A Benchmark for Metric Facial Depth Estimation and Its Evaluation},
  author    = {Marinescu, Alexandru I. and Borza, Diana L. and D{\u{a}}r{\u{a}}bant, Adrian S.},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations \& Datasets Track},
  year      = {2026}
}