Table 1.
The ablation study on the COCO dataset.
| Methods | Parameters (M) | FLOPs (G) | AP (%) | AP50 (%) | AP75 (%) |
|---|---|---|---|---|---|
| RetinaNet baseline | 37.74 | 95.66 | 32.5 | 50.9 | 34.8 |
| +ClassDecoder | 35.03 (−2.71) | 70.30 (−25.36) | 34.6 (+2.1) | 53.5 (+2.6) | 36.1 (+1.3) |
| +Global Extract Encoder | 36.95 (+1.92) | 90.45 (+20.15) | 36.2 (+1.6) | 55.7 (+2.2) | 37.8 (+1.7) |
| +Attention | 37.45 (+0.5) | 90.65 (+0.2) | 38.3 (+2.1) | 58.3 (+2.6) | 39.3 (+1.5) |
| +Anchor-free | 37.31 (−0.14) | 89.95 (−0.7) | 38.9 (+0.6) | 59.1 (+0.8) | 39.6 (+0.3) |
| +Data augmentation | 37.31 (+0) | 89.95 (+0) | 41.3 (+2.4) | 61.8 (+2.7) | 41.5 (+1.9) |