Dataset generator

Every time when you start to retrain a model you need a good dataset however often what you need does not exists. As a solution you can either create one manually, what takes a long time or you can generate it by using a dataset generator application. In this page we introduce you a sample dataset generator program, what is able to generate Pascal VOC style dataset with annotations of the bounding box coordinates. It supports also the creation of Darknet style dataset.

Dataset generator program

Let's suppose that you want to create a dataset with 5000 items. Each training data looks like the image at the left side, contains twelve cards with different symbols. A possible solution is to take 5000 pictures and annotate them manually one by one. Probably, it would be the most perfect dataset, however it would cost much time to create.

Example data with SET cards.

An easier way is to use an application for dataset creation. It will not be a perfect dataset because of the lower quality of the images, however it will be enough to retrain your model. It is a perfect solution when you don't have the final dataset, but you want to try out to train your newly created neural network.

With this data generator program you can generate 5000 items in a minute. You only need some raw data like backgrounds and cards.

Installation

Prerequirements:

Java JDK 1.8 or later;
Git version control system;

Compile and run the project:

Clone the Github repository from here: https://github.com/szaza/dataset-generator.git;
From the following links you can download some raw data:
- backgrounds;
- SET cards;
Save and extract them respectively into the /dataset/bacgrounds and /dataset/cards directories;
In the configuration file you are able to set the size of the generated images and also the number of the cards on the background;
Compile the application with Gradle by running the ./gradlew clean build -xtest command from the root directory;
Run the application with ./gradlew run;
As a result you should see the generated dataset under the /dataset/output directory;

Generated data with debug option

The generated directory structure under the /dataset/output directory should look like:

DarknetConf/ - contains the darknet configuration files
- voc.names - similar to the labels.txt.
- train.txt - it contains the paths for the images
- val.txt - it contains the paths for the images
- voc.data - data file for darknet framework
VOCDataSet/ - contains the Pascal VOC dataset
- Annotations/ - contains the xml style annotations
- JPEGImages/ - it contains the generated images
- labels/ - it generates the *.txt files for darknet
VOCLabels/
- labels.txt - label file needed to darkflow

The JPEGImages folder contains the generated images. The bounding boxes will be present on the generated data, if the debug flag is set.

Google Sites

Report abuse