✦
Summary
ORBIT is a benchmark for evaluating camera-pose estimation in difficult real-world video. It uses panoramic footage to create perspective-view test clips with recoverable reference trajectories, addressing the shortage of trustworthy evaluation data for scenes with motion, blur, crowds, or weak visual texture.
Key Features
The benchmark measures camera-pose estimation and structure-from-motion performance.
Full 360-degree context supports more robust reference camera tracking.
Perspective clips are cropped and reprojected from panoramic source footage.
Crowds and independently moving objects are among the evaluated challenges.
Low-light and low-texture scenes are included in the challenge categories.
High-speed camera motion is a dedicated evaluation category.
Examples compare methods including COLMAP, MegaSaM, and VGGT-Long.
The project links public code and the benchmark dataset.
About ORBIT
The dataset first estimates camera motion from the broader context of 360-degree video. Selected views are then cropped and reprojected into harder perspective clips. Methods are evaluated against the resulting reference trajectories, with examples and category-specific analyses showing where reconstruction systems struggle.
ORBIT is useful for researchers comparing traditional and learned structure-from-motion systems under realistic failure conditions. Public code, dataset links, and qualitative comparisons support inspection and reproduction. The project also uses the ORBIT++ name, while retaining ORBIT for the underlying benchmark clips and evaluations.