Closed muzeyen closed 2 months ago
Hi, thanks for your interest in our work.
The simplest way to realize the multi-class tracking is to tracking each class seperately. That is, you just need to prepare the MOT format files for each class, and no major code changes are required. Optionally, you can store the class id
to replace the -1
after the score
, and modify the codes to reject association between different classes.
For further questions about preparing detection files, maybe this file can help you: https://github.com/dyhBUPT/StrongSORT/blob/master/others/AuxiliaryTutorial.md
Thank you for your response. In the first suggestion, you mentioned producing output in a separate mot format for each class. What should these outputs be like? For example, I have a sequence and should I produce two separate txts for this sequence (sequence01), sequence01_person_mot.txt and sequence01_car_mot.txt?
Yes, you can try it.
Hello,
Firstly, I'd like to mention that I'm new to this field. I have detection outputs available in YOLO format (class_id, x_center, y_center, width, height, score).
Sample output:
The model is trained for 2 classes, which are car and person (0: person, 1: car). I am keeping these outputs saved in a text format for each frame.
Example:
frame1.txt:
frame2.txt:
To run StrongSORT using these outputs, if I understand correctly, I should first obtain a file in the MOT format. However, this format doesn't have a specific field for class_id in the format you mentioned, "frame_id, -1, x, y, w, h, score, -1, -1, -1". How should I produce an output in case of multiple classes and multiple detections? Should the -1 after frame_id represent the class_id? Which Python scripts should I use in what order to prepare my detection outputs for use in StrongSORT? Should I have which files when I want to run StrongSORT?
I apologize for the simplicity of my question. I would be grateful if you could explain the steps in this regard. Thank you in advance.
Regards.