I'm still toying a round with this and after a few iterations (up to a training set of around 10,000) it's generally working quite well - picking up various sizes and angles
Some strange false positives too though but the CoreML model I'm using isn't reporting the confidence levels correctly...