search

mikel-brostrom / boxmot

BoxMOT: pluggable SOTA tracking modules for segmentation, object detection and pose estimation models
GNU Affero General Public License v3.0
6.79k stars 1.72k forks source link

low metrics when eval #1758

Open dyhgo opened 1 day ago

dyhgo commented 1 day ago

Search before asking

Question

I execute the default command "python3 tracking/val.py --yolo-model yolov10x --reid-model osnet_x0_25_msmt17.pt --tracking-method botsort --verbose --source ./tracking/val_utils/MOT17/train". But the metric results are low, like {"HOTA": 25.661, "MOTA": 18.76, "IDF1": 22.929}. the problem seems to be that poor performance of yolov10x on MOT17. why is that ?

dyhgo commented 1 day ago

i use default yolov10x as detector, it perform good on MOT17-9 (big objects) but perform bad on MOT17-13 (many small objects)

mikel-brostrom commented 1 day ago

The results displayed under README are generated using the detections generated by a Yolox-X model, the same as in the bytetrack paper

dyhgo commented 10 hours ago

I read the paper, the detector is Yolox-X trained on CrowdHuman and half training set of MOT17. Do you have trained weights of Yolox-X

mikel-brostrom commented 5 hours ago

python tracking/val.py --benchmark MOT17 --yolo-model yolox_m.pth --tracking-method bytetrack --source tracking/val_utils/data/MOT17/train