Hello everyone,
I am running an object detection project on the Arduino VENTUNO Q using Edge Impulse and the Linux Runner, but I am having trouble running models exported for the Qualcomm QNN target.
Hardware and software
- Board: Arduino VENTUNO Q
- OS: Ubuntu Linux ARM64
- Edge Impulse Linux Runner: v1.25.3
- QAIRT libraries: 2.49.0
- Camera: Logitech C920
- Input resolution: 512 × 512 RGB
- Object classes:
Problem
Models V64, V65, V67, and V68 fail during inference with:
Classifying failed, error code was -3
The error occurs from the first frame and repeats continuously. The camera capture and image resizing pipeline work correctly.
Model V23 runs without this error but returns empty bounding boxes, so successful inference has not yet been confirmed.
During process inspection, V23 showed QNN/HTP-related libraries and FastRPC/DMA-BUF resources. These were not visible in the V64 process during the same type of inspection. Kernel logs did not show relevant QNN or HTP errors.
Questions
- Is there a recommended Edge Impulse export configuration for object detection on the VENTUNO Q?
- Are there known compatibility issues between the QNN model export, Linux Runner v1.25.3, and QAIRT 2.49.0?
- How can I obtain the underlying native error behind
-3?
Has anyone successfully deployed a custom Edge Impulse object detection model on the VENTUNO Q using QNN?
Thank you!