Fig. 1. Gauss heat map of central key points. (a) Example 1; (b) example 2; (c) example 3; (d) example 4
Fig. 2. CenterNet detection algorithm
Fig. 3. Structure of Hourglass-104
Fig. 4. Standard convolution and Pyramid convolution
Fig. 5. Pyramid convolution kernel structure and pyramid convolution residual block structure. (a) Shallow layer pyramid convolution;(b) shallow middle layer pyramid convolution; (c) middle layer pyramid convolution; (d) deep layer pyramid convolution; (e) pyramid convolution residual block
Fig. 6. Pyramid Hourglass-104 network structure
Fig. 7. Strip pooling module
Fig. 8. Strip pooling head. (a) Sharing scheme; (b) unshared scheme
Fig. 9. SIXray_OD dataset.(a) Example 1; (b) example 2; (c) example 3; (d) example 4
Fig. 10. Comparison of detection results. (a) CenterNet; (b) proposed algorithm
Category | Knife | Scissors | Wrench | Gun | Pliers | Total |
---|
Number of images | 12488 | 9272 | 31608 | 23488 | 18128 | 69744 |
|
Table 1. Statistics of SIXray_OD dataset
Network | Backbone | mAP50 /% |
---|
SSD | VGG16 | 71.89 | YOLOv3 | DarkNet-53 | 64.34 | Faster R-CNN | VGG16 | 78.41 | CenterNet | Hourglass-104 | 86.6 |
|
Table 2. Comparative experimental results of different detection networks
Experiment | Py_Hourglass_104 | Strip pooling | IoU loss | AP /% |
---|
mAP50 | Gun | Knife | Pliers | Wrench | Scissors |
---|
CenterNet | | | | 86.6 | 95.69 | 90.44 | 88.21 | 83.62 | 75.04 | Experiment 1 | ✓ | | | 87.3 | 96.00 | 91.23 | 88.65 | 84.53 | 76.09 | Experiment 2 | | ✓ | | 87.5 | 96.21 | 91.16 | 89.12 | 84.07 | 76.94 | Experiment 3 | | | ✓ | 87.4 | 95.86 | 90.64 | 89.54 | 85.72 | 75.25 | Experiment 4 | ✓ | ✓ | | 88.0 | 96.38 | 91.12 | 89.86 | 85.92 | 76.73 | Experiment 5 | ✓ | | ✓ | 87.7 | 96.12 | 91.53 | 89.05 | 85.18 | 76.62 | Experiment 6 | | ✓ | ✓ | 87.9 | 96.15 | 91.85 | 89.24 | 85.10 | 77.06 | Experiment 7 | ✓ | ✓ | ✓ | 88.3 | 96.40 | 91.88 | 89.90 | 85.90 | 77.43 |
|
Table 3. Ablation experimental results of each improvement point
Scheme | Pyramid convolution residual block used |
---|
Scheme 1 | [shallow、shallow middle、shallow middle,middle、deep] | Scheme 2 | [shallow、shallow、shallow middle、shallow middle、deep] | Scheme 3 | [shallow、shallow、shallow middle、middle、deep] |
|
Table 4. Pyramid convolution scheme
Experiment | Scheme 1 | Scheme 2 | Scheme 3 | mAP50 /% |
---|
CenterNet | | | | 86.6 | Experiment 1 | ✓ | | | 87.3 | Experiment 2 | | ✓ | | 87.0 | Experiment 3 | | | ✓ | 86.9 |
|
Table 5. Comparison results of different pyramid convolution schemes
Experiment | Da | D | mAP50 /% |
---|
CenterNet | | | 86.6 | Experiment 1 | ✓ | | 87.3 | Experiment 2 | | ✓ | 84.7 | Experiment 3 | ✓ | ✓ | 87.0 |
|
Table 6. Experimental results of pyramid convolution using position comparison
Experiment | Backbone | Preprocessing subnet | Strip pooling head 1 | Strip pooling head 2 | mAP50 /% |
---|
No | | | | | 86.6 | Experiment 1 | ✓ | | | | 85.6 | Experiment 2 | | ✓ | | | 86.4 | Experiment 3 | | | ✓ | | 87.4 | Experiment 4 | | | | ✓ | 87.8 |
|
Table 7. Experimental results of strip pooling using position comparison