data augmentation tool/feature

[mahdi-shafiei]

I was wondering what tools or features are available in caffe for typical data augmentation steps for image datasets?

[nicodjimenez]

I have the same problem. I don't want to re-create a dataset with ~20 million images every time I change my augmentation strategy, so I implemented on the fly augmentation. I implemented an augmentation step which randomly perturbs images by modifying data_layer.cpp. Every time datum.ParseFromString appears in this file, I add my own function Perturb(&datum), which modifies the datum by distorting the image and resizing it (I removed the requirement that all images in the leveldb have the same size because I don't want to distort downsampled images).

I think support for on the fly processing from the leveldb is sufficiently important to merit inclusion in caffe. Does anyone have any thoughts/suggestions on a clean way to implement this?

[bhack]

I think that augmentation can share many things with "synthetic" dataset generation

[bhack]

See also http://hal.inria.fr/hal-00979464/

[kloudkl]

Sander Dieleman implemented realtime augmentation as described in "My solution for the Galaxy Zoo challenge".

[bhack]

Another interesting experiment on augmentation: http://infoscience.epfl.ch/record/199706

[zimenglan-sysu]

try multi-scale？see this paper：http://arxiv.org/abs/1406.2984

[futurely]

@shelhamer, four existing implementations haven't been taken care of for a long time: #1386, #1865, #1900, #2252.

[longjon]

Please ask usage questions on the caffe-users list, and document related tools and techniques on the wiki. Thanks!

[bhack]

@senecaur Are you interested to contribute something back with a PR from your fork?

[futurely]

https://github.com/dmlc/mxnet/blob/master/src/io/image_augmenter.h