Many yolov8 models are trained on the VisDrone dataset. Office code https://github.com/ultralytics/ultralytics.
You can train according to my parameter settings or download my model. All model parameters were trained on a single A40 GPU. Using yolov8s as main comparison baseline.
| model | input size | other hyper-parameters | val mAP50 | val mAP50-95 | test mAP50 | test mAP50-95 | model & log |
|---|---|---|---|---|---|---|---|
| yolov8s | 640 | default | 0.404 | 0.24 | 0.32 | 0.184 | |
| yolov8l | 640 | default | 0.453 | 0.278 | 0.361 | 0.213 | |
| yolov8s | 960 | default | 0.496 | 0.308 | 0.396 | 0.234 | |
| yolov8s | 960 & mutil-scale | default | 0.492 | 0.304 | 0.399 | 0.236 | |
| yolov8s-p2 | 960 | default | 0.519 | 0.321 | 0.417 | 0.244 | |
| yolov8s-p2 | 960 | minxup=0.4 | 0.522 | 0.325 | 0.431 | 0.255 | |
| yolov8s-p2 | 960 | minxup=0.4, val iou=0.6 | 0.531 | 0.327 | 0.437 | 0.256 |
The parameters settings for best performance can be found in the default.yml file.
- copy_paste
- reduce image scale (+/- gain)