We are living in interesting times. In times like these, when personal health and fitness becomes a criticality, shelter-in-place poses new challenges. Currently, there are many Youtube videos on fitness that one could follow from their homes, but, there is no one to correct the person learning from the video if they make any incorrect move. It is usually a one-way communication. This leads to many injuries especially for beginners and it is also one of the deterring factors among people who want to learn a new fitness regime. Also, as a fitness expert, your outreach is often limited due to language limitations and distance. There are many yoga teachers in India who are among the best but their spoken language is Hindi which becomes a language barrier for people in the west. We envision a healthy society where any fitness expert can teach anyone without the barriers of language and distance with AI-powered application. Idea was to build an application that can bring such fitness experts to your home.

What it does

It's as simple as 1-2-3: 1)You simply log on to our app. 2) Choose the workout you want to do. 3) You then set your laptop on a table or on your bed and begin!  When you begin, the app guides you through a calibration step. After which, you can proceed to do your exercise. We have an extremely accurate auto rep counting feature, which counts your reps. You can see from the video it pauses when you pause. Once you’re done with the exercise, you can see aggregate statistics about your pose and form, get tips and log your workout to a leaderboard, and even download your video!

More importantly, our AI-powered fitness application will correct your posture if you make a wrong move while learning from the fitness expert in the video. As an added bonus, It will translate the video in your native language so that each and every instruction is crystal clear. Thus, having a personal trainer at your disposal. The application offers a list of fitness programs and their experts. A user gets to choose a program and a language of their choice. Each program has a set of lessons before you get a completion certificate. Each lesson has a video of the expert where he/she will teach an exercise/asana. Each video will be divided into steps so that it is easy for the user to follow. After each step, the user will be given some time to imitate the steps shown by the instructor. AI pose detector will closely monitor the moves of the student and correct the student if they have incorrect posture. It will try to match the student’s posture with the instructor and give you a thumbs up if you do it correctly

How we built it

Let’s talk a bit about the technical architecture. We have two aspects we are very proud of. Firstly getfit is a privacy first app. When you begin an exercise the model is downloaded to your local computer for evaluation. This is a variation of the posenet model that we have finetuned by running a few epochs on google cloud. All evaluation happens locally, using tensorflow.js So, none of the video is uploaded anywhere. Only your aggregates are.

Secondly, this is an entirely serverless app. All authentication and data is stored in google firebase’s realtime db. Each exercise just has a config. So, this is extremely extensible for all sorts of exercises. But we see it’s applications in a wide variety of domains like yoga, sports pose correction, and even physiotherapy.

Accomplishments that we're proud of

  1. Seamless end to end product where a beginner can start learning any fitness program with AI-powered personal trainer.
  2. Making the learning experience from videos more engaging and easier to understand.
  3. A business model for trainers. Any trainer can join our platform and have a global outreach by creating their fitness program on our application.
  4. Robust to the real-world - works across all camera angles, lighting's, and users.

What's next for GetFit

Go Live!

We believe that personalized interactive physical education is the future, and we know how prohibitively expensive it is. And we think getfit bridges the gap. Hope you enjoyed this, thank you!



firebase, google-posenet, javascript, semantic-ui, tensorflow

Devpost Software Identifier