Reproduce “LucidFusion: Reconstructing 3D Gaussians with Arbitrary 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
Hao He, Yixun Liang, Luozhou Wang, Yuanhao Cai, Xinli Xu, Hao-Xiang Guo, Xiang Wen, Yingcong Chen
LucidFusion reconstructs 3D Gaussian models and estimates camera poses from an arbitrary number of unposed, sparse input images, presenting a feed-forward solution to a common challenge in 3D reconstruction. The method demonstrates superior geometry and visual quality in sparse-view reconstruction, outperforming baselines on PSNR, SSIM, and LPIPS metrics, and provides faster, more accurate pose estimation (lower median rotation and translation errors) compared to existing feed-forward and optimization-based approaches.
Key insights
- Aligning pixel-wise 3D coordinates relative to a main camera view (RCM) effectively resolves orientation ambiguities common in world-coordinate-based multi-view methods.
- Interpreting pixel-wise relative coordinates as 3D Gaussian centers and supervising their differentiable rendering (RCG with 3D Gaussian Splatting) is critical for achieving global 3D consistency and improving reconstruction quality.
- Decoupling the learning of pixel-wise coordinate mappings (RCM) from global 3D consistency enforcement (RCG rendering) through a two-stage training scheme stabilizes the optimization and yields improved results.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.