Question
Need guidance for running YOLO-FastestV2 INT8 on STM32N6570-DK – Incorrect bounding boxes and false detections
Hi ST Community,
I am trying to deploy a YOLO-FastestV2 INT8 object detection model on the STM32N6570-DK. The model is successfully converted and runs on the board, but I am facing issues with the detection results.
Model Information
- Model: YOLO-FastestV2 (INT8)
- Number of classes: 6
- Parameters: 235,120
Model Input
Input_0_out_0
uint8 [1,1,320,320]Model Outputs
Quantize_297_out_0
int8 [1,12,20,20]
Quantize_313_out_0
int8 [1,3,20,20]
Quantize_311_out_0
int8 [1,6,20,20]
Quantize_330_out_0
int8 [1,12,10,10]
Quantize_346_out_0
int8 [1,3,10,10]
Quantize_344_out_0
int8 [1,6,10,10]Issues I am Facing
- The bounding boxes are not displayed at the correct positions on the camera image.
- The detected bounding boxes appear to be shifted or incorrectly scaled.
- Even when the camera is not viewing any object (or the camera feed is unclear), random bounding boxes continue to be displayed.
- These false detection remain on the screen even though there should be no valid detection.
Verification Performed
I also tested the quantized ONNX model on my PC using the same model.
The ONNX model produces correct detection results:
- Bounding boxes are in the correct positions.
- No false detection are observed when no object is present.
This suggests that the model itself is working correctly, and the issue is likely related to the STM32 deployment, post-processing, or output decoding.
Request for Guidance
I would appreciate guidance on the following:
- Is there an example of running a custom YOLO-FastestV2 model on STM32N6570-DK?
- Is there a reference implementation for decoding YOLO-FastestV2 outputs on STM32?
- Are there any requirements regarding:
- output tensor ordering,
- parameters,
- anchor configuration,
- confidence threshold,
- NMS implementation,
- or coordinate scaling?
- Could incorrect post-processing cause the bounding boxes to appear at incorrect positions and generate false detections?
- Are there any debugging steps you would recommend to verify whether the outputs are being decoded correctly on the device?
Attachments
I am attaching:
- Screenshot of the quantized ONNX model output on PC (correct detection).
- Screenshot of the STM32N6570-DK display showing the incorrect bounding boxes.
Any guidance or example project for deploying YOLO-FastestV2 on STM32N6570-DK would be greatly appreciated.
Thank you!
