Slightly off topic, but in the live demo it looks like model model wasn't really trained with images of a shovel that is diagonal in the frame, with the tip in top right and the base bottom left. The model detects all kinds of things but not the shovel - a spoon, a rifle, baseball bat etc