Additional sample code to run image classification and object detection on a Coral USB accelerator from a Raspberry Pi, filling some gaps in the official demos:
-
I would like to see object detection working on a video stream where rectangles are drawn on the preview to indicate object locations.
-
Because Pi Camera's frame width is restricted to multiples of 32 and height to multiples of 16, it cannot match some models' required input size. For example, if you tell Pi Camera to resize to Inception V3's required 299x299, it actually rounds up to 320x304. The captured image has to be cropped before passed to Edge TPU.
Resulting files are:
-
classify_capture.py: Original official demo, kept for reference. -
classify_capture_crop.py: This preemptively checks the model input size against Pi Camera's capable frame size. The image is cropped, if necessary, before passed to Edge TPU. -
detect_capture_crop.py: Use theAnnotatorclass borrowed from AIY Vision Kit to draw bounding boxes around detected objects. It also crops images if camera's frame size cannot match model input size.