Fig. 1. Structure of CSMS network
Fig. 2. Context extraction module. (a)Stacked model; (b) context-sensitive module
Fig. 3. Effective feature fusion structure
Fig. 4. Precision-recall curves on Wider Face validation set. (a) Easy; (b) Medium; (c) Hard
Fig. 5. Qualitative detection results
Model | Stackedmodel | CSMS context-sensitive module |
---|
Number ofparameters | 5*3×3 | 3*3×3 | Equivalentlayer | 3×3,5×5,7×7 | 3*3×3,3*5×5,7×7 |
|
Table 1. Comparison of context extraction module
Level | CSMS-sglDm | CSMS |
---|
Easy | 81.2 | 92.6 | Medium | 83.6 | 91.0 | Hard | 39.8 | 82.5 |
|
Table 2. Comparison of mAP for scale invariance design%
Level | Stacked model | CSMS |
---|
Easy | 92.5 | 92.6 | Medium | 90.8 | 91.0 | Hard | 82.0 | 82.5 |
|
Table 3. Comparison of mAP for context-sensitive module%
Level | OHEM | Focal Loss |
---|
Easy | 91.9 | 92.3 | Medium | 90.7 | 90.8 | Hard | 81.4 | 81.6 |
|
Table 4. Comparison of mAP for training methods focused on hard sample samples%
Level | CSMS-pl | CSMS-plDlt | CSMS |
---|
Easy | 92.3 | 92.6 | 92.6 | Medium | 90.8 | 90.8 | 91.0 | Hard | 82.0 | 82.2 | 82.5 |
|
Table 5. Comparison of mAP for feature fusion%
Method | Contribution | mAP / % |
---|
FocalLoss | Context-sensitivemodule | Effectivefeature fusion | Easy | Medium | Hard |
---|
Baseline | | | | 91.9 | 90.7 | 81.4 | CSMS | √ | √ | √ | 92.3 | 90.8 | 81.6 | | √ | √ | 92.4 | 90.8 | 82.2 | | | √ | 92.6 | 91.0 | 82.5 |
|
Table 6. Ablation study of CSMS on Wider Face test set
Method | mAP /% |
---|
Easy | Medium | Hard |
---|
Method inRef. [15] | 86.2 | 84.4 | 74.9 | Method inRef. [15] | 92.5 | 91.0 | 80.6 | Method inRef. [14] | 89.9 | 87.4 | 62.9 | Method inRef. [17] | 91.9 | 90.7 | 81.4 | Method inRef. [17] | 93.1 | 92.1 | 84.5 | CSMS(VGG-16) | 92.6 | 91.0 | 82.5 | CSMS(VGG-16)+Pyramid | 93.8 | 92.1 | 85.1 |
|
Table 7. Comparison of detection performance