New dataset integration

For this tutorial we are going to integrate the MOT20 dataset. Please refer to their documentation to download the dataset on your own machine. You can download all the data from this link.

TrackingDataset extension

The first step is to create a new class that extends the TrackingDataset class. This class will be used to load the data from the dataset and to provide the structure for the pipeline to work on. The TrackingDataset class is defined in tracklab/datastruct/tracking_dataset.py.

A TrackingDataset is an abstract class that allows to work on a subset of the dataset. It is composed of TrackingSet which represent the different splits of the dataset. These TrackingSet contain the information about ground truth detections, images and videos metadata which are stored as of pandas dataframes.

Basically, you need to here define the way to load the files and fill the pandas dataframes that respectively contain the information about the detection ground truths, images metadata and videos metadata for each split of your dataset. The minimum required information to be included in the pd.DataFrame are the following:

detections	image_metadatas	video_metadatas
[image_id, video_id, ..]	[video_id, file_path, frame, ..]	[name, nframes, ..]

/!\ do not forget to set the indexes of your dataframes to a unique ID. The update of the results accross the pipeline relies on the uniqueness of the indexes.

You can find the example of the MOT20 implementation of the TrackingDataset that we implemented in tracklab/wrappers/dataset/mot20.py. You will then need to add your new class to the tracklab/wrappers/dataset/__init__.py file.

Config file

Our framework works with the Hydra configuration system which takes advantage of a hierarchical configuration via files. This is very convenient to easily change modules.

Basically, you will need to add a .yaml file in configs/dataset which will be converted to a dictionary and will contain all the arguments required for your new class. The _target_ element indicates the object you want to instantiate for the creation of your dataset. It will point to our new class tracklab.wrappers.MOT20. You can find the configuration file in configs/dataset/mot20.yaml.

Then you have to change in the main config file (configs/config.yaml) the entry in datasets in defaults to mot20 (the name of the new file without the extension).