Research Article
OAB-YOLOv5: One-Anchor-Based YOLOv5 for Rotated Object Detection in Remote Sensing Images
Table 3
Comparisons between baseline method and the proposed YOLOv5m + OAB on DOTA-v1.5 test-dev (OBB task).
| OBB results | Methods | Params | PL | BD | BR | GTF | SV | LV | SH | TC | BC | ST | SBF | RA | HA | SP | HC | CC | mAP |
| YOLOv5m | 23.4 M | 74.79 | 74.80 | 46.99 | 55.73 | 66.42 | 76.44 | 88.41 | 86.78 | 73.53 | 80.65 | 40.85 | 65.30 | 69.88 | 76.44 | 48.56 | 18.15 | 65.23 | +OAB (without ViT) | 22.4 M | 79.68 | 77.2 | 46.89 | 53.63 | 66.29 | 75.67 | 88.49 | 90.54 | 76.03 | 78.28 | 46.65 | 62.79 | 67.90 | 77.50 | 45.78 | 22.60 | 65.81 | +OAB (with ViT) | 20.3 M | 80.67 | 77.2 | 46.37 | 60.13 | 66.08 | 75.80 | 87.85 | 90.64 | 78.12 | 79.54 | 46.28 | 70.58 | 67.70 | 75.42 | 53.63 | 28.15 | 68.02 |
|
|
OAB: one-anchor-based method; ViT: vision transformer.
|