以參考影格為基礎的無人機影片穩定:動態城市場景基準評測
Consensus Capture in Reference-Based UAV Video Stabilization: A Benchmark on Dynamic Urban Scenes
這項基準評測量度動態城市場景中無人機影片穩定的「共識捕獲」問題,涵蓋靜態及動態場景。未遮罩 ORB 在 304 組即時配對中有 28 組出現捕獲,角點誤差中位數為 18.8 px;遮罩則令所有管線均不再出現捕獲,且領先管線維持不變。研究同時發布基準數據及 Songdo Vision OBB,包含 5,419 段 4K 無人機影片中的 274,190 個定向框。
Reference-frame stabilization registers every frame of a hovering-drone video onto one reference frame, and its registration pipelines are usually picked on feature-matching benchmarks where only the camera moves. Over a busy intersection, vehicles moving together can pull the robust homography toward their own motion (consensus capture), while inlier counts and ratios still look normal.
We measure this with a two-part benchmark:
- Static: ten pipelines scored under exactly known homographies on 4K aerial frames, with tuning on a site-disjoint split.
- Dynamic: real hovers selected from the video alone, at an offline and a real-time budget, scored against a reference audited with manually annotated ground control points.
Unmasked ORB is captured on 28 of 304 real-time pairs; its captured estimates have a median corner error of 18.8 px with typical inlier statistics. Masking removes capture for every pipeline and leaves the leading pipelines unchanged.
Released: code (pip install stabilo-benchmark), the benchmark data with all results (Zenodo), and Songdo Vision OBB, with 274,190 oriented boxes on 5,419 4K drone frames.
來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co