Author Affiliations
1School of Mechanical Engineering, Inner Mongolia University of Science & Technology, Baotou , Inner Mongolia 014010, China2School of Information Engineering,Inner Mongolia University of Science & Technology, Baotou , Inner Mongolia 014010, China3Inner Mongolia University of Technology, Hohhot , Inner Mongolia 010051, Chinashow less
Fig. 1. Flow chart of CW algorithm
Fig. 2. Improved quadruplet network including two triplet networks
Fig. 3. Training the loss function of WIDER FACE data set
Fig. 4. Side face detection
Fig. 5. Small face detection
Fig. 6. CW algorithm clustering similar face nodes
Fig. 7. Clustering with nodes
Fig. 8. Isosurface display map
Fig. 9. Learning rate varying with epoch
Fig. 10. Total loss varying with epoch
Fig. 11. Accuracy varying with epoch
Fig. 12. Time varying with epoch
Fig. 13. ROC for different networks when subspace number is 5
Fig. 14. ROC for different networks when subspace number is 10
Fig. 15. ROC for different networks when subspace number is 20
Fig. 16. ROC comparison between different networks and FaceNet
Fig. 17. Multi camera face clustering based on Siam16
Fig. 18. Multi camera face tracking results of street view based on Siam16
Parameter | K-means | DBSCAN | HAC | MCL | CW |
---|
Algorithm complexity | O(n) | O(n2) | O(n3) | O(n2) | O(n)-O(n2) | Unknown number of clusters | × | × | √ | √ | √ | Real-time monitoring | × | × | × | × | √ |
|
Table 1. Performance comparison between CW algorithm and several clustering algorithms
Model | Feature size | Accuracy /% |
---|
1-DTNCNN16 | 1024 | 99.47 | 2-DTNCNN16 | 4096 | 99.51 | 3-MSMLCNN16 | 1024 | 99.09 | 4-MSMLCNN16 | 4096 | 99.21 | 5-TrHardCNN16 | 1024 | 99.02 | 6-TrHardCNN16 | 4096 | 99.11 | 7-QuadrupletCNN16 | 1024 | 98.45 | 8-QuadrupletCNN16 | 4096 | 98.85 | 9-TripletCNN16 | 1024 | 98.25 | 10-TripletCNN16 | 4096 | 98.78 | 11-SiamCNN16 | 1024 | 95.99 | 12-SiamCNN16 | 4096 | 96.21 | 13-ResNet50 | 4096 | 95.87 | 14-VGG-19 | 4096 | 94.14 | 15-VGG-16 | 4096 | 92.46 | 16-AlexNet | 4096 | 89.04 |
|
Table 2. Comparison of recognition rates of multiple detection networks on LFW data set
Tracker/Year | SR0.5 /% | SR0.75 /% | AO /% |
---|
LWL(*)[26]/2021 | 95.1 | 85.2 | 86.7 | LWL[27]/2020 | 92.4 | 82.2 | 84.6 | PrDiMP-50[28]/2020 | 89.6 | 72.8 | 77.8 | DiMP-50[29]/2019 | 88.7 | 68.8 | 75.3 | Siam16/2021 | 90.4 | 74.9 | 79.6 |
|
Table 3. Performance comparison on GOT-10k validation set