Frame-Consistent Feed-Forward 3D Gaussian Splatting from Sparse Unposed Images

Key points

• A natural instantiation pairs a frozen pose estimator, VGGT, with a frozen reconstruction backbone, TokenGS.1
• Through per-camera analysis across 107 scenes from two datasets, DL3DV and MipNeRF360, using COLMAP poses as the reference, we find that the rotation error after alignment is small (median 1.6 degrees), whereas the translation scale is offset by a large, dataset-consistent factor.1
• Controlled ablations over 140 DL3DV scenes show that the primary obstacles are two integration failures rather than pose error: a coordinate-frame inconsistency and a translation-scale mismatch, both of which can be corrected deterministically.1
• In the sparse setting, where reconstruction uses only four context images under the standard pose-free protocol, the same pipeline matches this all-view quality, yielding a fully COLMAP-free pipeline without any refinement step.1

Sources (1)

• [1]Frame-Consistent Feed-Forward 3D Gaussian Splatting from Sparse Unposed Images
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 16, 07:15 AM

  Feed-forward Gaussian models have made novel view synthesis almost instantaneous.

  A natural instantiation pairs a frozen pose estimator, VGGT, with a frozen reconstruction backbone, TokenGS.

Extractive summary: sentences quoted from the sources.