I'm try to generate a dataset used for object detection(about 1e6 pictures). I have images of objects, and backgrounds. But I think adding some filters on objects & objects+background is good for model training. After investigation, I found I have to use those following filters:
distortion, including shear with curve, distortion with sphere
cast lights on the img.
other simple filter like rotate, resize, blur, noisy, color gradation...
I wonder if there is any library, which is simple to use(input args + [w, h, 4], return [w, h, 4]), natural, fast, and most of all, it can add lights and lens light on img and distort it.
Or any library that can work with python during mxnet training.
After observing Photoshop, I have tried to make my own filters using sphere, ellipse and para-curve formulas by PIL, opencv and scimage, However, those filters are not natural and not as good as Photoshop does.
I also tried gimp-python, but I want to use it in a pure python program and I want to process the pictures during training. Maybe speed is really important.
These blog seems a practical way of distortion: http://paulbourke.net/miscellaneous/imagewarp/