FuguReport

ORBIT++: Benchmarking SfM in the Wild with 360° Video

Authors Sara Sabour, Linyi Jin, Richard Tucker, Amir Hertz, Marcus Brubaker, Saurabh Saxena, Junhwa Hur, Andrea Tagliasacchi, Deqing Sun, David J. Fleet, Richard Szeliski, Noah Snavely
Affiliations Google / Simon Fraser University / University of Toronto
Categories Evaluation / Benchmarking / Benchmark for SfM in real-world scenarios, Application / 3D Reconstruction / Camera pose estimation in 360° video, Task / Pose Estimation / SfM camera pose accuracy measurement
License CC BY 4.0

Abstract Overview

This paper presents ORBIT++, an updated version of the ORBIT benchmark for evaluating structure-from-motion and camera pose estimation on difficult real-world videos derived from online 360° footage. The benchmark is built by estimating pseudo-ground-truth trajectories from full panoramic videos using a custom rig-based SfM pipeline, then rendering challenging perspective clips with varied fields of view and rotation patterns. After filtering and validation, the dataset contains 308 clips drawn from 77 unique 360° videos, covering scenarios such as low texture, dark scenes, high-speed motion, crowds, egocentric captures, and fluid motion. The authors position the benchmark as a way to test methods in settings where existing synthetic or simpler datasets do not adequately reflect in-the-wild failure cases.

Novelty

The distinctive contribution is leveraging panoramic 360° web videos to recover reliable camera trajectories across a full omnidirectional field of view and then deriving difficult perspective-view benchmark clips from those same videos. This setup separates the omnidirectional information used to construct ground truth from the limited-view perspective inputs used during evaluation.

Results

Evaluations on ORBIT show that classical, optimization-based, and feed-forward methods all struggle on the benchmark, with every evaluated baseline exhibiting a failure rate above 36%. MegaSaM and RoMo+MegaSaM attain the lowest mean absolute trajectory errors (0.67 m and 0.59 m) while still failing on 39.09% and 38.56% of clips, whereas COLMAP fails on 56.49%, and ORB-SLAM2, VGGT-Long, and MonST3R fail on over 94% of clips under the evaluation threshold. The ground-truth generation pipeline achieves 0.07 ± 0.04 m ATE on the 360Loc dataset and 0.0 ± 0.0 m on Synthetic 360, confirming its high fidelity.

Key Points

  1. ORBIT++ constructs a challenging SfM benchmark from 360° web videos by estimating trajectories on the full panorama and then rendering difficult perspective clips with controlled rotations and fields of view.
  2. The benchmark contains 308 validated clips from 77 source videos and explicitly covers multiple real-world challenge categories, including low texture, dark lighting, high camera speeds, crowds, egocentric motion, and fluid scenes.
  3. Across evaluated baselines, no method is robust on ORBIT: even top-performing methods retain failure rates near 39%, with challenge analysis revealing distinct failure profiles between bundle-adjustment and feed-forward approaches.

References

This page was created using generative AI such as GPT-5, Claude Opus 4, Gemini 3, Gemini 3.1 Flash Image, and their higher-end successor versions. No guarantee can be made regarding its contents.