Hello community,
I am currently working on deploying a full YOLOv8n (80-class COCO) model on the STM32N6570-DK and integrating it with the Object Detection example from the STM32 AI Model Zoo Service.
I closely followed this tutorial:
How to deploy YOLOv8/YOLOv5 Object Detection models
and used the official quantization scripts here:
YOLOv8 Quantization Scripts
Model & Quantization Setup
Base model: yolov8n.pt (Ultralytics official)
Exported to: SavedModel format
Quantization: done per STM32AI model zoo tutorial
Input: uint8 (0–255)
Output tested as:
uint8 input / float output
uint8 input / int8 output
Case Input Type Output Type Result
(1) | uint8 | float | Incorrect inference — bounding boxes & class scores both wrong |
| (2) | uint8 | int8 | Bounding boxes coordinates always same (incorrect), classification correct |
After debugging case (2), I found:
The bounding box coordinates remain constant across frames and detections, even before post-processing.
The output tensors are quantized correctly, but the model’s raw outputs (prior to postprocess) seem fixed.
The default post-process used in the deployment script (deploy.py) is set for INT8 output models (POSTPROCESS_OD_YOLO_V8_UI). (Permalink)
When we changed this manually to POSTPROCESS_OD_YOLO_V8_UF (for float output) in the STM32N6 Object Detection app and reflashed, the results remained incorrect.
Thank you in advance for your help and guidance
