I haven’t published my own dataset because I believe that everybody’s use case will be slightly different (different room, different sources of heat, different expected distance to trigger the detection…).
However it’s quite easy to make your own dataset:
- Configure a platypush cron that takes IR pictures through the
camera.ir.mlx90640.capture action at regular intervals and let it run for a day or two. You should hopefully get a few hundreds/thousands B/W images that can be used to train the model.
- Use the
utils/label.py script from my
imgdetect-utils repo to label the images (it’s quite easy: create two folders,
positive, within your images folder, run the script and press “1” if an image is negative and “2” if it’s positive)
- Use my train notebook to train a Tensorflow model based on those images