Reproduce “FlexSplat: Flexible Feed-Forward 3D Gaussian Splatting without Point Cloud Correspondence” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Amir Sabbaghziarani, Hanting Ye, Maria Gorlatova, Yi Ding

FlexSplat introduces a feed-forward framework for 3D Gaussian Splatting that reconstructs compact, high-quality object models from uncalibrated multi-view images, predicting camera poses and depth without ground-truth. This approach achieves novel view synthesis performance comparable to state-of-the-art methods that require known camera parameters.

Key insights

  • Jointly training the geometry front-end with the Gaussian decoder allows the camera and depth predictions to co-adapt to the NVS task, making the system robust to initial geometry estimation errors.
  • Depth-guided Gaussian parameterization combined with multi-view deformable cross-attention enables the system to absorb moderate errors in self-estimated camera poses and depth by establishing a cross-view consensus for each Gaussian.
  • Employing a compact, resolution-independent Gaussian representation (decoupled from input image resolution and number of views) ensures efficiency without sacrificing reconstruction quality, using significantly fewer primitives than pixel-aligned methods.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.