Table 6.
Comparison of different models on the ROUD2023.
| Method | Model | Backbone | mAP@.5:.95% | mAP@.5% | mAP@.75% | AP-s% | AP-m% | AP-l% |
|---|---|---|---|---|---|---|---|---|
| One-stage | SSD30 | VGG16 | 43.4 | 73.4 | 45.4 | 11.7 | 31.6 | 48.4 |
| RetinaNet23 | ResNetXt101 | 50.7 | 79.3 | 54.5 | 14.3 | 39.2 | 56.1 | |
| FreeAnthor39 | ResNetXt101 | 55.0 | 82.4 | 59.8 | 17.0 | 42.9 | 60.7 | |
| NAS-FPN13 | ResNet50 | 51.4 | 78.9 | 55.2 | 14.4 | 38.3 | 56.7 | |
| ATSS40 | ResNet101 | 52.9 | 80.3 | 56.9 | 16.4 | 41.1 | 58.6 | |
| YOLOF45 | ResNet50 | 50.1 | 80.0 | 53.8 | 11.2 | 37.4 | 55.9 | |
| FocusDet(ours) | STCF-EANet | 62.2 | 84.8 | 64.3 | 17.1 | 44.2 | 63.4 | |
| Two-stage | Faster R-CNN8 | ResNetXt101 | 52.8 | 81.8 | 57.5 | 17.2 | 40.9 | 58.2 |
| Cascade R-CNN27 | ResNetXt101 | 54.8 | 81.1 | 59.7 | 16.8 | 42.2 | 60.6 | |
| Dynamic R-CNN36 | ResNet50 | 54.4 | 81.3 | 60.3 | 17.1 | 42.8 | 60.0 | |
| DetectoRS40 | ResNet50 | 57.8 | 83.6 | 63.6 | 20.4 | 45.0 | 63.7 | |
| Libra R-CNN43 | ResNetXt101 | 54.8 | 82.8 | 60.5 | 16.5 | 43.1 | 60.6 | |
| ThunderNet46 | ShuffleNetV2 | 41.7 | 67.9 | 44.6 | 8.8 | 25.6 | 46.7 | |
| Key-point based | Grid R-CNN47 | ResNetXt101 | 53.7 | 81.1 | 58.4 | 17.7 | 41.2 | 59.1 |
| RepPoints41 | ResNet101 | 55.4 | 83.7 | 60.4 | 17.7 | 43.3 | 60.8 | |
| CornerNet15 | HourglassNet | 41.9 | 60.3 | 43.7 | 9.5 | 33.2 | 43.7 | |
| Center-point based | FCOS48 | ResNetXt101 | 50.7 | 79.5 | 50.4 | 18.0 | 40.0 | 56.2 |
| FoveaBox49 | ResNet101 | 52.1 | 81.4 | 56.0 | 15.1 | 40.5 | 57.5 | |
| FSAF50 | ResNetXt101 | 48.7 | 78.5 | 51.2 | 15.7 | 38.0 | 53.9 | |
| Guided Anchoring42 | ResNetXt101 | 56.7 | 84.2 | 62.0 | 18.1 | 44.0 | 62.6 |
The bolded performance is the best one.