Caffe Image Database Prepare

When you download image datasets from the internet, they often come seperated into folders per class.

Caffe requires you to input the files as one text file, containing the path to the image and the class label (as a number).

This script creates such a file from an image database folder, generating class numbers.

It is even able to split up the images between training and test set by an arbitrary ratio.

Data

The directory structure of your database should be like this

Database
	|- Class_1
		|- Image_1
		|- Image_2
			...
			
	|- Class_2
		|- Image_1
		|- Image_2
		|- Image_3
			...
	...

Usage

The command line interface is simple

caffe_image_prep <database directory> [options]

This will generate a file called output.txt containing the path to every image together with its class label.

Options

Option	Short Version	Value	Default	Meaning
--train	-t	string	output.txt	The output file for the training set
--test	-T	string	none	The output file for the test set
--table	-ta	string	none	File listing the class tables along with the classnames
--random	-ra	none	none	If set, the images order is randomly shuffled, resulting in different training and test sets each time
--ratio	-r	float	0.8	The ration to between training and test set images. Must be between 0 and 1.
--verbose	-v	int	1	The amount of verbosity. 0 = silent, 1 = output summary (default), 2 = processing output

Examples

The following will store the path, along with the class label for every image in the 101_ObjectCategories directoy into `output.txt``

caffe_image_prep 101_ObjectCategories

The following will split the images into a training and a test set by a ration of 0.75 (75% traning, 25% test). The output will be stored in the files data_training.txt and data_test.txt repectively.

caffe_image_prep 101_ObjectCategories --train data_train.txt --test data_test.txt --ration 0.75

The following will output both a training and a test set with the default split ration of 0.8. Furthermore it will create a table file data_table.txt listing the class labels (numbers between 0 and n) together with the class names (derived from the foldernames).

caffe_image_prep 101_ObjectCategories --train data_train.txt --test data_test.txt --table data_table.txt

Install

Download the repository or clone it using

git clone https://github.com/00SteinsGate00/CaffeImageDataPrepare.git

Copy the script to some place inside your $PATH, for example /usr/bin and set the correct permissions.

sudo cp caffe_image_prepare.py /usr/bin/caffe_image_prep
sudo chmod +x /usr/bin/caffe_image_prep

Licence

MIT Licence