Reproduce “FullCircle: Effortless 3D Reconstruction from Casual 360$^\circ$ Captures” 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
Yalda Foroutan, Ipek Oztas, Daniel Rebain, Aysegul Dundar, Kwang Moo Yi, Lily Goli, Andrea Tagliasacchi
FullCircle masks operators in dual-fisheye video and reconstructs scenes directly from raw views. It uses pretrained vision models to detect and segment the person, then applies those masks in pose estimation and reconstruction. Across nine casual-capture scenes, it averages 29.6 dB PSNR versus 28.3 for SLS-MLP, measured against clean tripod references; this result is limited to the reported novel-view test.
Key insights
- Raw dual-fisheye inputs offer superior camera calibration robustness and higher-quality novel view synthesis compared to perspective or undistorted images, especially for out-of-distribution viewpoints.
- Leveraging pre-trained vision foundation models with a targeted re-centering approach effectively and automatically masks the camera operator in distorted 360° imagery without requiring task-specific training.
- Directly training radiance fields on raw fisheye images, rather than relying on undistortion, is more effective for high-fidelity 3D reconstruction from 360° captures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.