Question Regarding Performance Gap Between Custom and ST Pretrained YOLOv8 Models

[VicChang]

Dear ST team,

I have followed the official tutorial:
How to deploy YOLOv8/YOLOv5 object detection models
and successfully built the deployment pipeline using Ultralytics yolov8n.pt.

However, I’ve noticed a significant performance gap between my converted model and the pre-optimized model provided by ST, in terms of inference speed under the same environment.

Am I missing any optimization steps during the export process?
Are there any recommended configurations or parameters to match the official model’s performance?

Thank you for your support.

[Laurent FOLLIOT]

Hello,
On which platform do you run the Yolov8n?
Regards,
Laurent

[VicChang]

Dear Laurent,

I conducted a performance comparison between the YOLOv8n model provided by ST and a model I converted myself using the official Ultralytics pre-trained YOLOv8n, with a focus on inference latency evaluation.

No retraining or modification was performed — I used the standard Ultralytics YOLOv8n model as-is.
The model was quantized and evaluated following the official ST tutorial:

https://github.com/STMicroelectronics/stm32ai-modelzoo-services/blob/main/object_detection/deployment/doc/tuto/How_to_deploy_yolov8_yolov5_object_detection.md

The evaluation was performed on the STM32MP257F-EV1 platform:

https://www.st.com/en/microcontrollers-microprocessors/stm32mp2-series.html

If you have any suggestions regarding the quantization process, deployment settings, or expected performance, I would greatly appreciate your input.

Thank you very much!