Artificial Intelligence Can Be Used For Many Things

As artificial intelligence and machine learning technologies improve and become more mainstream, we are seeing more web developers using them. Pixel Hop, a British digital design agency based in Brighton, recently shared a cool example of AI that uses the pose detection library maintained by Google at TensorFlow.JS. The project included a bit of Halloween whimsy, and it basically used facial feature detection in order to turn visitors with cameras into skeletons.

According to the developer, the project involved the WebGL back-end libraries of TensorFlow, an API for video recording, another API for audio and, and yet another for sharing reactions on social media apps. The project runs surprisingly smooth considering all the server calls and Vue libraries queried, but it is pretty neat nonetheless.

As you can imagine, pose detection is not an easy ML task. Think about all the possible variations in the poses that humans strike at any moment. Google developers at the TensorFlow project have been using computer vision to overcome these difficulties; you can see through the aforementioned Halloween skeleton project that they are getting better, but we still quite a few years of R&D left on this specific ML task.

Many AI researchers have long believed that a major advance in pose estimation can come from a combination of sophisticated deep-learning technology and large training data, but it hasn’t really happened. In the most famous example, last year Google released a paper that claimed its DeepFace model — essentially an artificially intelligent image classifier — was able to recognize a new face within a single second, using a dataset of 10 million face images. But when DeepFace is pointed at just a random patch of the image, the performance is still nowhere near that good — DeepFace’s accuracy is between 85 and 90 percent on the average image.

Microsoft is trying to prove that pose estimation can be as accurate as object recognition. The tech giant is using its advanced neural network developed for object recognition as an effort to shift it over to pose and facial recognition. In the near future, we may be able to see more pose detection routines incorporated into digital user interfaces. For more information click here http://trick-or-treat.pixelhop.io/.