An improved Tiny YOLOv3 for real-time object detection

Figures & data

Figure 1. The feature fusion of the 26 feature map.

Figure 2. The maxpooling and concatenation of the two scale.

Figure 3. The specific implementation of the maxpooling and concatenation.

Figure 4. The output feature map of the 13 × 13 feature scale.

Figure 5. The network structure of the improved Tiny YOLOv3.

Table 1. Different network model size.

Algorithm name	Module size/M
YOLOv3	246.5
Tiny YOLOv3	33.2
Improved TinyYOLOv3	55.9

Figure 6. The CIoU loss function for bounding box regression.

Table 2. Detection speed of different algorithms.

Algorithm name	GPU fps/(frames·${\text{s}}^{-1}$)	CPU fps/(frames·${\text{s}}^{-1}$)
Tiny YOLOv3	35.5	1.2
Improved Tiny YOLOv3	32.5	1.1

Figure 7. The setscrew AP of Tiny YOLOv3.

Figure 8. The nut AP of Tiny YOLOv3.

Figure 9. The setscrew AP of improved Tiny YOLOv3.

Figure 10. The nut AP of improved Tiny YOLOv3.

Table 3. The average precision of the different classification.

Algorithm name	Class	AP (%)
Tiny YOLOv3	Setscrew	90.76
Tiny YOLOv3	Nut	32.18
Improved Tiny YOLOv3	Setscrew	93.88
Improved Tiny YOLOv3	Nut	97.24

Figure 11. The mAP of Tiny YOLOv3.

Figure 12. The mAP of improved Tiny YOLOv3.

Figure 13. The comparison diagram of experimental results.