Spaces:

SkalskiP

/

YOLO-ARENA

like 30

Running on CPU Upgrade

@SkalskiP Thanks for your fantastic demo and detailed evaluation! We are the team of YOLOv10 and we appreciate your efforts very much. However, we find some inconsistent inference results of YOLOv10 in this demo and our local environment. We are not sure where the reason lies in, so we want to seek your help and guidance. Specifically, we use the same YOLOv10m model in this demo to infer the images in our local environment. The detailed process is below

wget https://skalskip-yolo-arena.hf.space/file=/tmp/gradio/f878616a10625ce7dba02bcb34df2df279273666/image.png
yolo export model=yolov10m.pt format=onnx opset=13 simplify half=True device=0
yolo predict model=yolov10m.onnx source=vehicles.png half conf=0.4

The inference result is below

However, we observe that the result is different from that of this demo.

Moreover, we want to clarify that due to that YOLOv10 adopts the different training strategy with YOLOv8 and YOLOv9, it may have different favorable confidence threshold to detect objects. Therefore, comparing the results of YOLOv10 with those of YOLOv8 and YOLOv9 on detecting small objects may be less appropriate with the same threshold. Besides, we note that different thresholds will not have the adverse impact on the inference latency of YOLOv10 because it does not rely on NMS. Therefore, we can freely set a smaller threshold for YOLOv10 to detect more small objects. For example,

wget https://skalskip-yolo-arena.hf.space/file=/tmp/gradio/048ecf37df927d6d7ed5528076f4133147f797d2/people-walking.png
yolo export model=yolov10m.pt format=onnx opset=13 simplify half=True device=0
yolo predict model=yolov10m.onnx source=people-walking.png half conf=0.05

The inference result is below

Thank you!