Datasets for ModelPack

ModelPack implements training iterators capable of reading Darknet format. Darknet format represents bounding box annotations in the following way:

class x-center y-center width height

where the class is the id of the class within the list of classes, x-center and y-center represent the normalized coordinate of the center point of the object concerning the image dimension, and the width, and height are the normalized dimensions of the object.

Annotations are saved in *.txt files where each line represents a different object on the input image. This means that a given image with N objects will contain a designated annotation file with N lines. The annotation file will have exactly the same name as its corresponding image but, with *.txt extension.

For more explanation on dataset format follow Darknet Dataset Schema for ModelPack and coco128 format. Download Playingcards dataset for testing purposes.

The training dataset will be stored in the following way:

dataset.yaml
dataset
|__images
     |__train
        |__00001.jpg
        |__00002.jpg
        ...
     |__validate
        |__07891.jpg
        |__07892.jpg
        ...
|__labels
     |__train
        |__00001.txt
        |__00002.txt
        ...
     |__validate
        |__07891.txt
        |__07892.txt
        ...

The dataset.yaml file contains the dataset description which includes the path to the training and validation sets as well as the name of the classes. For Playingcards dataset, the yaml file looks like this: