[HOW-TO] Specify classes to be detected by IMX500 AI Camera

[kustusch]

Is it possible to specify or filter certain classes of objects to be detected?

I would like to modify demo script:
https://github.com/raspberrypi/picamera2/blob/main/examples/imx500/imx500_object_detection_demo.py

Per my understanding, models based on COCO dataset are able to detect around 80 classes of objects listed here:
https://github.com/raspberrypi/picamera2/blob/main/examples/imx500/assets/coco_labels.txt

In my project I would like to narrow down number of classes detected for specific model used (eg. YOLOx).

Such capability is implemented by Ultralytics. The model.predict() accepts "classes" argument to filter classes to be detected.

Is it possible to accomplish something similar with IMX500?

[davidplowman]

Hi and thanks for the question.

I recall that Ultralytics and Sony collaborated recently on a feature which allows you to export Ultralytics models in Sony IMX500 format. Might this be what you need? So far as I recall however, the IMX500 is only currently supporting YOLO v8 models.

I think I tried this out for myself at some point in the past. There were quite a lot of scary-looking warnings that came out during the expoort, but the model worked fine at the end.