Mirror Mirror: Crowdsourcing Better Portraits

摘要

We describe a method for providing feedback on portrait expressions,rnand for selecting the most attractive expressions from largernvideo/photo collections. We capture a video of a subject’s facernwhile they are engaged in a task designed to elicit a range of positivernemotions. We then use crowdsourcing to score the capturedrnexpressions for their attractiveness. We use these scores to train arnmodel that can automatically predict attractiveness of different expressionsrnof a given person.We also train a cross-subject model thatrnevaluates portrait attractiveness of novel subjects and show how itrncan be used to automatically mine attractive photos from personalrnphoto collections. Furthermore, we show how, with a little bit ($5-rnworth) of extra crowdsourcing, we can substantially improve therncross-subject model by ”fine-tuning” it to a new individual usingrnactive learning. Finally, we demonstrate a training app that helpsrnpeople learn how to mimic their best expressions.