Reproduce “OF$^3$GS: On-the-Fly Feed-Forward 3D Gaussian Splatting from Unposed Images” 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
Ruiyang Chen, Feiran Li, Chu Zhou, Zonglin Li, Zhanyu Ma, Heng Guo
The OF3GS framework introduces a feed-forward method for on-the-fly 3D Gaussian Splatting from unposed images, overcoming the latency of optimization-based techniques and the geometric instability of prior feed-forward approaches. It achieves high-quality novel view synthesis by stabilizing intrinsic camera parameters and refining Gaussian primitive placement under causal constraints, demonstrating competitive performance with offline methods in sparse-view settings and robust generalization.
Key insights
- Stabilizing camera intrinsics via a decoupled head, rather than predicting them frame-wise, is critical for maintaining long-term coordinate consistency and preventing severe geometric drift in causal, unposed online reconstruction.
- Rigid unprojection of depth to 3D Gaussian primitives is insufficient in causal online settings; introducing dynamic offsets provides necessary spatial flexibility to correct for accumulating pose-depth errors.
- Asymmetric novel-view-weighted supervision during training encourages the network to learn robust 3D geometry rather than overfitting to input-view appearance, improving generalization and novel-view consistency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.