search

fullfanta / CNN_based_Face_Alignment

Face Alignment using Convolutional Neural Networks
20 stars 11 forks source link

Did you do Data Augmentation? #1

Closed wuqiangch closed 7 years ago

wuqiangch commented 7 years ago

How many train images do you get after do Data Augmentation?

fullfanta commented 7 years ago

@wuqiangch

Before augmentation, I collected 3283 images from IBUG(135 images), AFW (337 images), LFPW (811 images), and HELEN (2000 images) of 300W. I augmented 10 times from each image so that I trained 32830 images.

wuqiangch commented 7 years ago

@fullfanta you only use the landmarks to get face ,and dont use face detector?

fullfanta commented 7 years ago

@wuqiangch

At the training time, I didn't use face detector. I tried opencv's face detector to get bounding box, but the result is very wrong. So I used only landmarks with scale, rotation, and translation augmentation.

But at the test time, I used opencv's face detector like mxnet/predict.py: https://github.com/fullfanta/CNN_based_Face_Alignment/blob/master/mxnet/predict.py#L17-L33