Train yolo on mot dataset about yolo_tracking HOT 7 CLOSED

For a SOTA people detector I would recommend you to follow: https://github.com/ifzhang/ByteTrack#data-preparation. WItht eh prepared data you can then easily use Yolov5 or 7 for training

from yolo_tracking.

Hi Mike I have ran the mot to coco format conversion. I now have the annotations folder containing some .json files. What should I do next? do I need to convert the dataset format from coco to yolo?

from yolo_tracking.

Yes, YOLO format is needed 100%. For all the people datasets

from yolo_tracking.

Hi Mike, I've just managed to start the training meanwhile I noticed that the box loss and obj loss has downward trend and the mAP sometimes goes up and down but always stuck around 0.74 with the highest currently attained at epoch 2 with 0.77 but higher box and obj loss. Is there anything to be made out of this?

Edit: the mAP fluctuation might be caused by using a smaller batch size.

from yolo_tracking.

Did you managed a successful training @rslim97 ?

from yolo_tracking.

@mikel-brostrom yes, I managed to train yolov7 for 300 epochs (around 26 hrs). The final weights converged to only 0.729 mAP. The best ckpt has about 0.748 mAP. I've tried on my dataset and the last weights performed worse (a lot of obvious missed detections), however sadly the best weights also just has a very slight improvement. I think the model has overfitted to mot dataset and unfortunately is not very suitable for my use case (has more close range view of people).

from yolo_tracking.

Add more datasources to your training data so that your model generalizes better: MOT17, MOT20, CrowdHuman, Cityperson, ETHZ

from yolo_tracking.