InfiniteRep: An open-source synthetic dataset for remote fitness and PT applications
If you build machine learning (ML) models for fitness or physical therapy (PT) applications, you know that off-the-shelf pose estimation models (like BlazePose) do not work well in production. There are simply too many poses and environmental variables that the model fails on.
We know because we tried. Here at Infinity, we’ve spent our entire careers as engineers building ML models that make sense of the world. When a fitness company asked us to build vision-based rep counting models, the first place we started was pose estimation. The existing models worked well enough on common poses, but broke down on some of our target exercises, especially when variables such as poor lighting, diverse body types, or occlusions came into play.
We needed to improve our pose estimation accuracy, but it was a pain to collect and label our own video data. We could either do our own custom data collection, or try to work post-processing magic on the poor off-the-shelf pose estimation results. Instead, we started a company to help all the ML engineers out there who, like us, were tired of not having the right data for the task at hand.
Infinity is a synthetic-data-as-a-service company. We synthesize the exact data that engineers need (with perfect labels) at the click of a button.
We’re proud to release the InfiniteRep dataset as a callback to the problem that got us hooked on synthetic data: remote fitness. InfiniteRep looks just like the videos collected by remote fitness and physical therapy companies, except it has perfect labels and purposefully-designed variation.
We hope InfiniteRep saves you innumerable hours and makes it possible to count squats, lunges, and crunches like never before.
If you need data that’s a little bit different (different movements, labels, demographics, or anything else), just drop me a line (lina@toinfinty.ai).
I read every email.
Download the Dataset
Download the InfiniteRep dataset here.
Github resources (README, demo notebooks, etc.) here.
Designed to look like your real-world data.
The InfiniteRep dataset contains 1,000 videos of diverse avatars doing multiple reps of 10 common exercises. It includes massive variation in the environment, lighting conditions, avatar demographics, and movement trajectories. From cadence to kinematic trajectory, each rep is done slightly differently (just like real humans).
It features:
- 1,000 videos (5–30 seconds each)
- 10 exercises
- Lifelike rep behavior (no two reps are done in the same way)
- 7 indoor home or gym scenes
- Diverse lighting conditions
- Varied demographics, including body shape, skin tones, and clothing
- 18 label and annotation types
Lifelike rep behavior
Each rep is done slightly differently. Just like a real human.
From cadence variation to varying kinematic trajectories, no two reps are done exactly the same way. We built in vast amounts of movement variation to enable you to train robust algorithms.
Choose from common exercises
The dataset includes 10 of the most common fitness and physical therapy (PT) exercises.
Or import any movement.
Import any movement into simulation and generate synthetic videos for it. Leverage your own motion capture (Mo-Cap) data or tell us the name of the exercise and we’ll take care of the rest.
Infinite avatars and environments
Varied rooms, camera angles, and lighting conditions.
The dataset includes 7 unique 3D scenes of home workout environments with a range of lighting conditions and camera angles.
Diverse avatars for equitable and fair ML.
Give everyone a great home fitness experience by designing for a range of skin tones, body types, and clothing. The dataset has a normal distribution of body-types and equal distribution of skin tones.
Open-source
Build without limits.
InfiniteRep is one of the few human-centric datasets that is available for both academic and commercial use.
Detailed labels
Pixel-perfect labels and rich annotations.
The dataset allows engineers to build rep counting models, pose estimation models, form-correction models, segmentation models, activity classification models, and much more. We are especially excited about the rep count labels since getting these gold-standard labels is notoriously difficult.
We’re excited to hear from you.
We’re a small team of engineers who built this dataset hoping it’d be useful to people like you!
Has the dataset been useful? We love hearing good things!
Has the dataset not helped your model? We’d love to help debug.
Do you want a different dataset? We specialize in generating bespoke synthetic data.
Drop me a line at lina@toinfinity.ai (I read every email).
Download the Dataset
Download the InfiniteRep dataset here.
Github resources (README, demo notebooks, etc.) here.
Infinity AI. Bespoke ML-ready datasets.
Infinity AI is a synthetic data-as-a-service company. We generate the exact training data you need to solve your ML problem. It’s as easy as picking your scene ingredients (people, movement, environment) and telling us whether you want a lot or a little variation of each. Get in touch (info@toinfinity.ai).
