Author Affiliations
School of Internet of Things Engineering, Jiangnan University, Wuxi, Jiangsu 214122, Chinashow less
Fig. 1. Schematic of optimized regional fully convolutional neural network model
Fig. 2. Connection modes of different convolution layers. (a) Traditional VGG convolution connection; (b) convolutional connection of 50-layer residual network
Fig. 3. Schematics of deformable convolution with different migration modes. (a) General convolution; (b) deformable convolution; (c) convolutional ideal arrangement; (d) convolutional rotation transformation
Fig. 4. Flow chart of deformable convolution
Fig. 5. Schematic of RPN network structure in model
Fig. 6. Pooling operation of ROI
Fig. 7. Pooling of deformable position-sensitive ROI
Fig. 8. Pooling of deformable position-sensitive ROI in the proposed method
Fig. 9. Small scale measurement in dark light
Fig. 10. Test under dark-light and background interference
Fig. 11. Multi-object boundary box overlapping test
Fig. 12. Multi-object and multi-scale test
Fig. 13. Small-scale occlusion test
Fig. 14. Multi-object occlusion test
Fig. 15. Two-object occlusion test
Fig. 16. Small-scale fuzzy object test
Fig. 17. Multi-object frontal head test
Fig. 18. Multi-object occlusion side head test
Method | Anchor | S-NMS | Deformable conv | Iterations | Test speed /(frame·s-1) | mAP /% |
---|
DPM Face R-CNN | - | - | - | - | - | 37.4 67.1 | Local-RCNN | - | - | - | - | - | 72.7 | Faster RCNN(ZF) | - | - | - | 50000 | 17.24 | 73.48 | Faster RCNN(ZF) | √ | - | - | 50000 | 17.24 | 74.50 | Faster RCNN(VGG16) | - | - | - | 50000 | 6.36 | 79.17 | Faster RCNN(VGG16) | √ | - | - | 50000 | 6.36 | 80.02 | R-FCN(ResNet-50) | - | - | - | 30000 | 7.29 | 81.00 | R-FCN(ResNet-50) | - | √ | - | 30000 | 7.29 | 81.12 | R-FCN(ResNet-50) | √ | - | - | 30000 | 7.29 | 81.96 | R-FCN(ResNet-50) | √ | - | - | 40000 | 7.29 | 82.00 | R-FCN(ResNet-50) | √ | √ | - | 40000 | 7.29 | 82.41 | R-FCN(ResNet-50) | √ | - | √ | 30000 | 7.29 | 82.76 | R-FCN(ResNet-50) | √ | - | √+ | 30000 | 6.95 | 83.24 | R-FCN(ResNet-50)* | √ | - | √ | 30000 | 7.04 | 82.83 | R-FCN(ResNet-101) | - | - | - | 30000 | 8.50 | 84.49 | R-FCN(ResNet-152) | - | - | - | 30000 | 8.32 | 84.32 | Note: “*”: position sensitive ROI align; “+” : deformable position sensitive ROI; “√”: corresponding network added. |
|
Table 1. mAP and test speed of different models on HollywoodHeads