Table 4 Double-Head R-CNN performance on the DeepPCB dataset.
Method | Input | mAP | \(AP_{s}\) | \(AP_{m}\) | mAR | \(AR_{s}\) | \(AR_{m}\) |
|---|---|---|---|---|---|---|---|
Double-Head R-CNN w/ Resnet50 (baseline3a) | \(640 \times 640\) | 74.7 | 75.1 | 74.8 | 79.2 | 79.4 | 79.1 |
Double-Head R-CNN w/ Resnet50 (baseline3b) | \(800 \times 800\) | 76.7 | 77.0 | 76.7 | 81.2 | 81.2 | 81.0 |
Double-Head R-CNN w/ Resnet101 (baseline4a) | \(640 \times 640\) | 74.5 | 74.7 | 74.7 | 79.2 | 79.0 | 79.3 |
Double-Head R-CNN w/ Resnet101 (baseline4b) | \(800 \times 800\) | 76.1 | 75.9 | 76.2 | 80.9 | 81.2 | 80.9 |
Double-Head R-CNN w/Resnet50+ACASEM | \(640 \times 640\) | 76.7 | 76.7 | 76.9 | 81.2 | 81.6 | 81.6 |
Double-Head R-CNN w/Resnet50+ACASEM | \(800 \times 800\) | 78.5 | 78.9 | 78.5 | 82.8 | 83.3 | 82.6 |
Double-Head R-CNN w/Resnet101+ACASEM | \(640 \times 640\) | 76.4 | 76.4 | 76.6 | 81.0 | 81.3 | 80.8 |
Double-Head R-CNN w/Resnet101+ACASEM | \(800 \times 800\) | 78.0 | 77.9 | 78.2 | 82.5 | 82.7 | 82.3 |
