Degerium Yolov11m_pose vs Yolov8m_pose

General Discussion
,
1

I am comparing the fps at batch size=8 between the yolov11m pose model downloaded from the Degerium Model Zoo and the yolov8m pose model already included in the hailo-apps-infra folder and am getting unexpected results.

I receive the expected 70+fps using the yolov8m pose .hef but am receiving <45fps when using the yolov11m pose .hef.

I expect higher fps when using the yolov11m pose model compared to the yolov8m given the efficiency gains for the yolov11m pose.

I compared both .hef files and they seem to be roughly the same size and fairly similar. Could this be explained with a difference in compilation flags? I have tried to compile and quantize the yolov11m pose model with Windows WSL but have had many issues. I also do not have access to a computer running Linux at the moment.

Thank you for your help!

2

Hi @Julian_Narvaez

For Hailo devices, it is not true that yolo11(n/s/m) models are faster than yolo8(n/s/m) (at least in our experience). Please see hailo_model_zoo/docs/public_models/HAILO8/HAILO8_object_detection.rst at master · hailo-ai/hailo_model_zoo for object detection benchmarks and you will observe that yolo8(n/s/m) models are always faster than yolo11(n/s/m) models. The same trend applies to pose and segmentation models.

3

Understood, thank you very much for your reply! Out of curiosity, why is that case?

4

Hi @Julian_Narvaez

Not all hardware options behave the same way. So, the gains seen by yolo11 on some hardware do not necessarily apply to all hardware options.