Feedback Welcome — What Would Improve These Datasets?

#1
by dxgl - opened

We're releasing this first collection (10 CC0 Polyhaven objects, 196×1024 sphere coverage) to get feedback from the NeRF/3DGS community before scaling up.

We'd appreciate your thoughts on:

Depth maps

  • Are the 8-bit and 16-bit depth maps accurate at object edges?
  • Is the near/far plane range useful, or would you prefer raw metric depth?

Masks

  • How do the binary masks handle transparent or reflective materials (glass, chrome)?
  • Any edge artifacts?

View distribution

  • We use a Fibonacci golden-angle spiral on a full sphere (±89° elevation). Does this distribution work well for your training method?
  • Would you prefer a different sampling strategy?

Point clouds

  • Are ~200k points sufficient for initialization?
  • Is the alignment with RGB views correct?

Missing modalities

  • Anything you need that's not included? (e.g., optical flow, semantic labels, metric depth)

Object selection

  • What categories should the next collection cover?
  • Any specific objects or material types you'd like to see?

Pre-trained splats

  • Are the splatfacto PLYs useful as baselines?
  • Should we include checkpoints or just the exported PLY?

All datasets are CC0 — use them however you want. If you find issues or have suggestions, please reply here or open a separate Discussion.

Interactive splat viewer: https://dx.gl/splat/index.html
Collection page: https://dx.gl/datasets/polyhaven-10
Pipeline details: https://dx.gl/for-research

dxgl pinned discussion

Sign up or log in to comment