vision

RPi Bot - Vision

Object Detection

The object detection code uses Neural Networks to identify objects in a given image or frame of a video.

The currently used model is MobileNet V3 (large), an SSD model (Single Shot object Detection) trained on the COCO dataset.

The choice of model is configurable to allow easier integration of future or alternate models, which enables testing and benchmarking of multiple models to determine a best fit for your computational requirements in a resource-constrained environment.

Testing / Benchmarking Object Detection

The object detection code can be run standalone for testing/benchmarking purposes. Simply call the script directly:

python3 object_detection.py

The script uses reasonable defaults, but configurable parameters are available. Check the help documentation:

python3 object_detection.py -h

A typical command during a non-desktop session (e.g., SSH) would be:

python3 object_detection.py --image image.jpg --benchmark

A typical command during a desktop session would be:

python3 object_detection.py --image image.jpg --save --show --benchmark

If you --save or --show images, the bounding boxes and labels are baked onto the images. If you also pass --benchmark, the neural network inference time is baked onto the image.