Yolov5 running with no detections on an OAK D Lite

jdmaunder

Hi,
I am trying to run Yolov5 (on a customized dataset) on an OAK D -Lite on a rasberrypi4, on the following two scripts from the below github repo:
main.py and main_api.py
https://github.com/luxonis/depthai-experiments/tree/master/gen2-yolo/device-decoding

I have used the posted google collab notebook and the following tool (http://tools.luxonis.com/) to convert my pytorch weights (pt file) a to Blob and Json file (files attached) but neither produced a bounding box detection. I was wondering if you could provide me some insight on what might the issue might be.

I am pretty confident its a configuration issue as other pre models from the depthai-python github repo run without an issue.

Thanks in advance,

erik

Hi jdmaunder ,
Have you selected the Yolov5 model in tools.luxonis.com? Could you also try lowering the confidence threshold in the json file? And feel free to share the .zip provided by tools.luxonis, so we can try it locally as well.
Thanks, Erik

jdmaunder

I figured out what my issue is, thank you for your help and the quick response.

erik

jdmaunder great! And feel free to share what the issue was, maybe it's relevant to other folks as well🙂

jdmaunder

For sure. One part of the issue was distance from my camera (i had to be further way than i thought) and background noise was messing up my inference when I was testing my video feed. The second part of the issue is Yolov5 itself. I am not to sure what the issue is here since I had significantly better results on the same dataset when I used Yolov3tiny instead of Yolov5 nano.