Reproduce “Wild3R: Feed-Forward 3D Gaussian Splatting from Unconstrained Sparse Photo Collection” 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
Yuto Furutani, Takashi Otonari, Kaede Shiohara, Toshihiko Yamasaki
Wild3R introduces a feed-forward 3D Gaussian Splatting method that produces appearance-consistent and transient-free 3D reconstructions from unconstrained sparse photo collections in a single pass. The approach achieves competitive perceptual quality (e.g., LPIPS 0.506 on Photo Tourism for 16 views) while being orders of magnitude faster (0.95s) than optimization-based methods, addressing challenges like varied illumination and transient objects.
Key insights
- Training on a synthetic dataset specifically engineered with diverse illumination and transient objects, like WildCity, is crucial for feed-forward 3DGS models to generalize effectively to unconstrained real-world photo collections.
- Explicitly enforcing appearance consistency conditioned on a reference view and supervising with transient-free depth maps enables robust disentanglement of static scene geometry from dynamic elements and varying lighting.
- Partial fine-tuning of pretrained components, rather than full retraining, efficiently adapts existing feed-forward 3DGS architectures for 'in-the-wild' conditions while preserving established capabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.