Presentation is loading. Please wait.

Presentation is loading. Please wait.

A Convolutional Neural Network Cascade For Face Detection

Similar presentations


Presentation on theme: "A Convolutional Neural Network Cascade For Face Detection"β€” Presentation transcript:

1 A Convolutional Neural Network Cascade For Face Detection
CV lab Chanmi you

2 A Convolutional Neural Network Cascade For Face Detection
H. Li, Z. Lin, X. Shen, J. Brandt, G. Hua, A Convolutional Neural Network Cascade for Face Detection, CVPR 2015

3 Contents Testing Process Training Process Experiments Conclusion

4 Testing process

5 12-net 12-calibration-net 24-net 24-calibration-net 48-calibration-net
Apply 12-net to obtain face/non-face classification Calibrate each patches +NMS Input 24-net Calibrated patches Extract 12x12 patch from whole image Face patches after 12-net Apply 24-net Face patches after 24-net +NMS 24-calibration-net 48-calibration-net 48-net Calibrate each patches Calibrate each patches +NMS Apply 48-net Face patches after 48-net Calibrated patches Output

6 12-net (Detection net) : 12x12 detection window
For each detection windows,

7 12-calibration-net CNN after 12-net for bounding box calibration
Given detection window (π‘₯, 𝑦, 𝑀, β„Ž) Calibration patterns (𝑁=45) For apply

8 12-calibration-net (cont’d)
Take the average results of the patterns After 12-calibration-net, Non-maximum suppression applied

9 24-net (Detection net) For multi-resolution, Fully-connected layer from 12-net is concatenated

10 24-calibration-net Similar with 12-calibration-net After 24-calibration-net, Non-maximum suppression applied

11 48-net (Detection net) For multi-resolution, Fully-connected layer from 24-net is concatenated. Relatively more complicated. After 48-net, Non-maximum suppression applied

12 48-calibration-net Relatively more complicated.

13 Training process Calibration nets Detection nets
For the 𝑛-th pattern [ 𝑠 𝑛 , π‘₯ 𝑛 , 𝑦 𝑛 ], apply [ 1/𝑠 𝑛 , βˆ’π‘₯ 𝑛 , βˆ’π‘¦ 𝑛 ] Detection nets 12-net Resize all training faces into 12x12 Randomly sample 200,000 non-face patches from background images Choose a threshold 𝑇 1 at 99 % recall rate 24-net Resize all training faces into 24x24 Densely scan all background images All detection windows with confidence score (after 12-net) larger than 𝑇 1 become negative training samples Choose a threshold 𝑇 2 at 97 % recall rate 48-net Resize all training faces into 48x48 Following same procedure. Neg: 5800 background images Pos: Annotated Facial Landmarks in the Wild(AFLW dataset)

14 Experiments Annotated Faces in the Wild (AFW dataset)

15 Experiments (cont’d) Face Detection Data Set and Benchmark (FDDB dataset)

16 Conclusion Face detection using CNN and cascade
Reject non-face regions quickly at low resolution Process accurate detection at higher resolution Calibration nets are introduced in the cascade to accelerate detection and improve bounding box quality


Download ppt "A Convolutional Neural Network Cascade For Face Detection"

Similar presentations


Ads by Google