ITP: Image Transformation Pursuit

The ITP algorithm is designed to select a small set of transformation to apply to an image classification dataset, to maximize the gain in accuracy

This code is a reimplementation of the one used for the following paper:

Transformation Pursuit for Image Classification Mattis Paulin, Jerome Revaud, Zaid Harchaoui, Florent Perronnin and Cordelia Schmid, CVPR 2014.

Dependencies

To make this code work, you need:

Python
Numpy
PIL
JSGD (Follow details on the dedicated webpage to install properly).

Sources

Code

The current version of the code is beta. It contains code to

Transform images
Select transformations

Test files

We provide the deep architecture descriptors used in our paper to test the program. They are extracted on the Caltech University Birds dataset with the DeCAF code available online. We use the fc7 layer, with the ReLU output. They are organized one transformed dataset after an other: taking the 5994 (resp 5794) first columns of the train (resp. test) gives the original untransformed dataset. In addition, we propose the 40 transformations used in our paper:

horizontal flip
10 crops
8 homographies
3 levels of jpeg compressions
5 scales
3 colorimetric transformations
10 rotations

The file is available here.