v1.0.0

First public release of Stabilo Benchmark, the benchmark behind the paper "Consensus Capture in Reference-Based UAV Video Stabilization: A Benchmark on Dynamic Urban Scenes" (preprint to follow).

It evaluates the complete registration pipeline of reference-based UAV video stabilization (feature detector, descriptor, matcher, robust estimator and object masking) on static and dynamic urban scenes, and scores how well each estimate aligns the static ground, not how many inliers it has.

Highlights

• Part I, static protocol: pipelines scored under exactly known synthetic homographies, with per-pipeline tuning on a site-disjoint split at a matched resolution and feature budget (static).
• Part II, dynamic protocol: real hovering-drone footage, cases chosen from the recordings by survey, built by prepare, scored against a pseudo-ground-truth reference (oracle) that is audited with manually annotated ground control points (annotate-gcp, audit), at an offline and a real-time budget (dynamic).
• Consensus capture made observable: registration error decomposed along the dominant motion of the annotated objects, and a capture fraction for how far an estimate follows the largest group of objects that moved together.
• Object masking as a controlled axis: none, horizontal boxes (HBB) and oriented boxes (OBB), applied before feature detection to every pipeline alike.
• Every number from a command: fetch downloads and verifies the paper's data record, and report regenerates its tables (LaTeX and CSV) and figures from the stored campaigns.
• A committed sample (three Part I scenes and one Part II case with its ground control points) runs both protocols without the full data, videos, detector weights or a GPU.

Install

pip install stabilo-benchmark

Python 3.11 to 3.13. To run the paper's configurations and the sample, clone the repository and pip install -e . (see the README).

Data

The paper's data record, 10.5281/zenodo.23623540 (CC BY 4.0), holds both parts' inputs, the surveys, every campaign's results and a copy of configs/songdo/:

stabilo-benchmark fetch songdo

Requirements

Built on Stabilo ≥ 1.4.5 for the registration pipelines and HBB2OBB ≥ 1.12.0 for the object annotation. HBB2OBB runs at the configuration selected on the human-verified oriented boxes of Songdo Vision OBB, released separately with the paper.