Table 8.
Accuracy performance with augmented dataset (30,000 images).
| Model | Backbone | Image Size | AP | AP50 | AP75 | APS | APM | APL |
|---|---|---|---|---|---|---|---|---|
| YOLOv3 [48] | Darknet-53 | 608 × 608 | 39.6% | 63.1% | 40.3% | 24.7% | 40.3% | 44.6% |
| YOLOv3-tiny [48] | Darknet-53 | 21.7% | 40.6% | 23.5% | 11.4% | 24.7% | 29.5% | |
| RFBNet [53] | VGG-16 | 40.3% | 61.9% | 42.6% | 23.1% | 43.4% | 49.7% | |
| SSD [54] | VGG-16 | 36.8% | 56.5% | 37.1% | 18.3% | 37.8% | 45.2% | |
| RefineDet [55] | VGG-16 | 41.6% | 61.4% | 41.9% | 22.6% | 43.2% | 45.8% | |
| YOLOv4 [17] | CSPDarknet-53 | 47.2% | 70.3% | 52.4% | 31.8% | 50.1% | 56.4% | |
| YOLOv4-tiny [17] | CSPDarknet-53 | 26.4% | 48.4% | 29.6% | 15.2% | 29.3% | 34.5% | |
| Improved YOLOv4 | CSPDarknet-53 | 48.6% | 72.6% | 53.7% | 31.3% | 51.5% | 58.1% |