search

yinguobing / cnn-facial-landmark

Training code for facial landmark detection based on deep convolutional neural network.
MIT License
623 stars 183 forks source link

Accuracy #87

Open wangping1408 opened 4 years ago

wangping1408 commented 4 years ago

Hello Is there a new evaluation indicator? Is there only the root mean square error used in the verification, and is the accuracy of the NME evaluation test used?

yinguobing commented 4 years ago

I think this project needs a more formal way of evaluation. Could you please introduce a little more about NME?

wangping1408 commented 4 years ago

NME normalized mean error is empolyed to measure the accuracy in many landmark detection paper. which averages normalized errors over all annoted landmarks.

yinguobing commented 4 years ago

Hi @wangping1408,

I have not found any matching built in metric from TensorFlow's official metrics. Maybe we need to write it by ourselves.

There is a page about adding custom metrics in keras. But I'm not sure if TensorFlow support it. You can find it here: https://keras.io/metrics/#custom-metrics