kfpandayang / data_augmentation_caffe

common data augmentation methods, implemented in caffe.

common data augmentation methods in caffe-based deep learning framework.

Data augmentation has shown great vitalness in boosting image based deep learning approach. I implemented more than 7 data augmentation methods based on caffe deep learning framework, including light correlation, rotation, random crop, blurring, color casting, vignetting, fish eye distortion. etc. Note that my implemetation is just an online data augmentation process, which means that, given an input image, I randomly choose to apply none/one/multiple transfomation listed above. See jitter_image_total( cv::Mat& cv_img_, const int crop_size ) function to detailed illustration.

#example images after data augmentation

From left to right: Original RGB image, fish-eye distortion, vignetting, three color casting (red, green, blue channel, respectively.). I did not given other augmentation results because they are easy to imagine. #Miscellaneous 1. I write this code when I was working at [Dress-Plus](http://www.dress-plus.com); 2. I only tested this code in caffe but it can easily be transfered to other deep learning framework. 3. I plan to add more data augmentation methods later on.