Reproduce performance of YOLOv9

[johannes-tum]

Hi,

to what extent could you already reproduce the results of the official YOLOv9 (and YOLOv7) repository?

Best wishes

Johannes

[tloki]

In my experience, Yolo-V9-S with coco trained weights (and used on subset of classes - 23 of them) exceeds accuracy of original implementation of yolov9 repo (also stock coco weights and 23 classes).

I tested usin pycocotools.

[johannes-tum]

That sounds great. Have you also tried larger models like the c model? Did you make any changes to the repository? Did you use the COCO pretrained weights from the official yolov9 repo? And what was the dataset you finetuned / tested on?

I am surprised that you even exceeded the performance, because this new repository deactivates some strong data augmentations like Mixup?