Author Affiliations
1Key Laboratory of Intelligent Medical Image Analysis and Precise Diagnosis of Guizhou Province, College of Computer Science and Technology, Guizhou University, Guiyang , Guizhou 550025, China2School of Information Engineering, Peking University Shenzhen Graduate School, Shenzhen , Guangdong 518055, China3Big Data Science and Intelligent Engineering Research Institute, Guizhou Education University, Guiyang , Guizhou 550018, Chinashow less
Fig. 1. YOLOv3 network structure diagram
Fig. 2. DenseNet network structure diagram
Fig. 3. DenseBlock internal structure diagram
Fig. 4. Structure of RD-Net feature extraction network
Fig. 5. MDC structure diagram
Fig. 6. First MDC2 channel connection diagram
Fig. 7. Complete network structure diagram
Fig. 8. PR curves of ablation experiments on two datasets. (a) Brainwash dataset; (b) HollywoodHeads dataset
Fig. 9. Comparison of PR curves between RDM-YOLOv3 and other methods on Brainwash dataset
Fig. 10. Comparison of PR curves between RDM-YOLOv3 and other methods on HollywoodHeads dataset
Fig. 11. Comparison of test results of two head datasets. (a) HollywoodHeads dataset; (b) Brainwash dataset
Network | D-CBL of DenseBlock1 | D-CBL of DenseBlock2 |
---|
Structure | Conv (1×1×32) | Conv (1×1×64) | BN | BN | Leaky-ReLU | Leaky-ReLU | Conv (3×3×64) | Conv (3×3×128) | BN | BN | Leaky-ReLU | Leaky-ReLU |
|
Table 1. Internal channel information of transport layer
Method | AP (RIOU=0.5) | FPS |
---|
HollywoodHeads | Brainwash |
---|
DarkNet-53 (baseline) | 0.689 | 0.752 | 19 | RD-Net | 0.750 | 0.801 | 28 | RD-Net+MDC1 | 0.782 | 0.846 | 25 | RD-Net+MDC1+MDC2 | 0.823 | 0.889 | 21 | RD-Net+MDC1+2×MDC2 | 0.868 | 0.931 | 16 |
|
Table 2. Comparison of ablation experiments on two data sets
Method | Backbone | AP (RIOU=0.5) |
---|
SSD | VGG16 | 0.568 | FCHD | VGG16 | 0.700 | YOLOv3 | DarkNet-53 | 0.768 | E2PD | GoogLeNet+LSTM | 0.821 | FRCN | VGG16 | 0.878 | HeadNet | ResNet-101 | 0.910 | RDM-YOLOv3 | RD-Net+MDC | 0.931 |
|
Table 3. Comparison results on Brainwash dataset
Method | Backbone | AP (RIOU=0.5) |
---|
DPM-Face | - | 0.370 | SSD | VGG16 | 0.621 | YOLOv3 | DarkNet-53 | 0.689 | FRCN | VGG16 | 0.712 | FCHD | VGG16 | 0.743 | TKD | LSTM | 0.750 | RDM-YOLOv3 | RD-Net+MDC | 0.868 |
|
Table 4. Comparison results on HollywoodHeads dataset