Building Damage Detection on xBD (test)
64.2PrecisionS + Aug
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| S + Augbackbone=YOLOX-S, architecture=siamese processing, augmentation=post-disaster shift augmentation during training2026.05 | 64.2 | 58.7 | 61.2 | 60.7 | |
| S + Abackbone=YOLOX-S, architecture=siamese processing, attention=cross-attention on D4–D52026.05 | 62.4 | 54 | 57.5 | 57.1 | |
| YOLOv12 S + A + Augbackbone=YOLOv12, architecture=siamese processing, attention=cross-attention on D4–D5, augmentation=post-disaster shift augmentation during training2026.05 | 62.3 | 54.4 | 58.1 | 60 | |
| Sbackbone=YOLOX-S, architecture=siamese processing2026.05 | 61.1 | 57.8 | 59.1 | 58.2 | |
| S + A + Aug + Comp(8)backbone=YOLOX-S, architecture=siamese processing, attention=cross-attention on D4–D5, augmentation=post-disaster shift augmentation during training, channel-wise latent compression ratio=82026.05 | 60.9 | 55.4 | 57.8 | 57.6 | |
| S + A + Aug + Comp(64)backbone=YOLOX-S, architecture=siamese processing, attention=cross-attention on D4–D5, augmentation=post-disaster shift augmentation during training, channel-wise latent compression ratio=642026.05 | 59.8 | 55.6 | 57.4 | 57 | |
| S + A + Augbackbone=YOLOX-S, architecture=siamese processing, attention=cross-attention on D4–D5, augmentation=post-disaster shift augmentation during training2026.05 | 58.9 | 57.4 | 57.9 | 57.2 | |
| EF (6ch)backbone=YOLOX-S, fusion=early fusion2026.05 | 58.3 | 57 | 57.4 | 57.4 | |
| S + A + Aug + Comp(64)−D3backbone=YOLOX-S, architecture=siamese processing, attention=cross-attention on D4–D5, augmentation=post-disaster shift augmentation during training, channel-wise latent compression ratio=64, feature level=removal of the finest-resolution feature level (D3)2026.05 | 56.9 | 48.6 | 52.3 | 49.7 | |
| S + A+D3backbone=YOLOX-S, architecture=siamese processing, attention=cross-attention extended to D32026.05 | 54.6 | 49.9 | 51.9 | 48.3 | |
| xView2 winneradaptation=Seg -> Det2026.05 | 50.4 | 52.1 | 51.2 | 38.8 |