torch-INRIA
makeClasses
is a simple script that loads face, torso, body and background samples from INRIA dataset and saves them in a ImageNet compatible format. This means, object are fairly centred and the sample size is 256 × 256 px².
In this first release, the dataset has to be downloaded manually from here and extracted into ~/Work/Datasets
.
In the code there is my user name (
~
=/Users/atcold
). I will probably update it in a following commit.
What this script does is simply extract every face, torso and body from the positive and background from the negative folder, cropping them square and saving them into a Extracted-data
folder, under a sub-directory representing their label.
Face
: are sized 1/3 of the width of the bounding box;Torso
: upper biggest square croppable from the bounding box;Body
: square with side equal to the hight of the bounding box. If the limits of the image are exceeded (the image is too narrow), the first/last column of pixel is replicated to fill the gap;Bckg
: squares of 128 px are cropped randomly from the negative samples.
And this is how they look like
How to run it
After changing the folder pointers in the code (which I will make soon automatic), in the src
directory, run
th makeClasses.lua
How is this done?
To have an idea of the information provided with the dataset, I wrote checkBox
which gives a general idea of what we can do with the data. To run it, go to the src
directory and type
qlua checkBox.lua