I'm impressed by the 3D effect on the labelled images. Not just the fact that they move around when you hover them, but the depth effect between the image and its labels.
If it were real, you could put a camera on the dog's head, and automatically shout at him if he starts chewing on non-approved items. The dog learns to behave and the AI learns from seeing more dog-POV data, creating some kind of adversarial Pet-NNet training.