This is the kind of cool work I want to see in computer vision research. Predictive models that do the right thing instead of more image classification tasks.