Master of Science in Analytics

Master of Science in Analytics

Deep Learning Yoga Pose Classification Capstone Project

Master of Science in Analytics students built an app that uses a one-step neural network to examine images of yoga poses and recognize the poses in order to provide feedback to the app's yoga-practicing user.


Video Transcript

More and more people are starting to practice yoga and everyone from beginners to experienced yogis needs instructor feedback. But, this feedback isn’t readily available outside a yoga studio.

So, students at UChicago are testing and applying neural network image recognition models to power an app that provides feedback to help aspiring yogis improve their practice anywhere they like.

Their app needed a model that could identify poses accurately, quickly, and from multiple camera angles. The team extensively tested two deep learning approaches to image recognition to see if either approach would fulfill their requirements.

The team prepared over seventy eight hundred images to train, validate, and test their models. The winning model would need to be able recognize eight poses, including standing, sitting, and inverted positions.

The team first tested a two-step process consisting of a pose-extracting neural network feeding into classification models. The team trained their image recognition model to recognize joints and sent this information to the second step: a classification model that matched the image to a yoga pose.

Even after significant testing and refinement, the two step model was only able identify one pose. So, they turned to one-step neural networks.

The team’s one-step neural network classifier provided impressive results. The team was able to successfully classify all eight poses with a one-step neural network based on a residual learning framework. 

Now, the team has an application of deep learning that can take an image of a yoga pose and return the name of the pose. From here, they’re making steps towards training their models with even more poses and expand their efforts to make meaningful yoga guidance and feedback accessible to more people.