Table 3 Faster R-CNN-FPN, YOLOv5 and YOLOv7 results for each coccinellids type in the test subset. Results are reported in terms of AP obtained with the best model according to the development subset. The Faster R-CNN-FPN (Faster) models are using ResNet-50 (R50) and ResNet-101 (R101) as base models, with IoU and GIoU as the loss, respectively. The best results for a coccinellid type using models within a family are highlighted with bold, while the best result overall is also shown in italics.
From: Detecting common coccinellids found in sorghum using deep learning models
Model | Coccinella septempunctata | Coleomegilla maculata | Cycloneda sanguinea | Harmonia axyridis | Hippodamia convergens | Olla v-nigrum | Scymninae |
|---|---|---|---|---|---|---|---|
Faster-R50-IoU | 67.1 | 62.7 | 60.3 | 65.2 | 55.1 | 67.0 | 63.0 |
Faster-R101-IoU | 70.0 | 63.9 | 58.8 | 70.7 | 54.5 | 69.8 | 65.1 |
Faster-R50-GIoU | 66.2 | 64.5 | 59.6 | 68.0 | 55.6 | 67.4 | 63.2 |
Faster-R101-GIoU | 70.1 | 66.6 | 59.7 | 68.8 | 58.3 | 69.0 | 67.0 |
YOLOv5n | 72.1 | 70.5 | 64.0 | 71.0 | 60.9 | 68.8 | 67.2 |
YOLOv5s | 76.3 | 71.8 | 68.8 | 72.7 | 63.9 | 73.1 | 69.3 |
YOLOv5m | 77.2 | 72.8 | 69.0 | 73.1 | 68.4 | 77.0 | 72.3 |
YOLOv5l | 77.1 | 73.3 | 70.1 | 74.2 | 69.2 | 75.3 | 72.0 |
YOLOv5x | 76.8 | 74.9 | 70.3 | 76.8 | 69.1 | 77.6 | 72.8 |
YOLOv7 | 80.4 | 76.4 | 72.2 | 76.2 | 70.5 | 75.5 | 71.0 |
YOLOv7-tiny | 75.5 | 68.8 | 63.7 | 72.7 | 65.3 | 70.3 | 61.4 |
YOLOv7-x | 74.8 | 69.8 | 65.4 | 73.1 | 62.8 | 69.4 | 63.1 |
YOLOv7-d6 | 74.2 | 66.4 | 61.1 | 72.4 | 62.8 | 69.8 | 50.4 |
