A mechanism to regularize the classifier layer by estimating the marginalized effect of label-dropout during training.
Rethinking the Inception Architecture for Computer Vision
Thanks to all the contributors of caffe and Rethinking the Inception Architecture for Computer Vision
There are 7 types of data augmentation (shift, zoomImg, rotateImg, modHSV, modRGB, JpegCompression , smoothFilter)
And shift、zoomImg and rotateImg are so important due to the unstability of face alignment.
- You should carefully read the code in src/caffe/layers/image_data_transform.hpp before you use those data augmentation.
- JpegCompression and smoothFilter harm the performance in face recognition in my experiments。
- Take care of parameters in image_data_transform.hpp(line 13 to 17) because those settings may be just suitable for the face images of my experiments(112X96 images same as sphereface)