search

xiangrufan / keras-mtcnn

mtcnn face detection transplanted to keras
MIT License
123 stars 45 forks source link

Landmarks #1

Closed Zumbalamambo closed 7 years ago

Zumbalamambo commented 7 years ago

How do I get landmarks as in this

xiangrufan commented 7 years ago

The landmark is commented out in previous version of code. I just fixed this problem and you should be able to see the landmark in this version.

Zumbalamambo commented 7 years ago

Thank you so much. May i know why is that it is slow in realtime? is there any ways to speed it up in cpu?

xiangrufan commented 7 years ago

This conv-net is very sensitive to resolution of input image, reducing resolution may be a good way to speed it up. The first layer process the input image at full resolution, which is very slow when the image size is large. it also gives much more positive detections to the second layer of mtcnn when resolution is large hence significantly slowed the detection process.

Zumbalamambo commented 7 years ago

May I know how do I train the model with reduced resolution?

xiangrufan commented 7 years ago

you do not need to retrain a new model. just use input video or image of lower resolution.