First, download and extract the training dataset and v3 training annotations from DETRAC.

Please note that the conversion process can take a while and requires a decent amount of free disk space. With my default settings on the DETRAC training set, I had 163,171 cropped images of vehicle sequences (which was only about 330MB) and 81,446 copied images for YOLO (which, including annotations, was about 5.13GB).

I followed this paper as a guideline for parameters. They removed all sequences of vehicles that had a truncation or occlusion threshold higher than 0.5. Those that then had less than 100 occurrences were ignored. Each vehicle in each satisfactory image is then cropped and resized to 100x100.

I've replicated that in the script so that you can choose your own thresholds and output image size. The script basically goes through the annotations and creates a dictionary of what to crop and what to ignore.

These are the default settings that I used for my own model:

• Occlusion threshold = 0.5
• Trucation threshold = 0.5
• Number of occurrences = 100
• Image size = (100, 100)

You can change these within crop_dataset.py or you can pass the arguments in the command line (except for image size).

For example,

crop_dataset.py --occlusion_threshold=0.6 --truncation_threshold=0.6 --occurrences=50

The cropped sequences should all be outputted to the DETRAC_cropped directory in Market 1501 format. Then, follow the instructions on how to train a Market 1501 dataset in the njojke's cosine metric learning repository.

I trained a YOLO v4 model using AlexeyAB's darknet. Follow the instructions there to train your own model.

I decided to include the occlusion and truncation thresholds here for the same reasons as when I trained the Deep SORT model. However, there's no image resize since there's no need to crop anything and no occurrence thresholds since YOLO isn't for tracking.

Like crop_dataset.py, you can change the default settings within the file or pass the arguments in the command line.

For example,

detrac_to_yolo.py --occlusion_threshold=0.6 --truncation_threshold=0.6

The script works in a similar way to crop_dataset.py by iterating through each annotation file and calculating the truncation and occlusion ratios of each object. Those that satisfy the requirements are counted with respect to its frame number. Frames that contain no qualifying objects in the end are then ignored.

The images are then copied to the DETRAC_YOLO_training directory and the corresponding annotations for each image are also calculated and outputted to DETRAC_YOLO_annotations.

A class list text file is also created for the outputted training images under detrac_classes.txt. You need this for training in Darknet and also for when you use the YOLO model for object detection.

I've also included the config file that I used when I was training with Darknet.