Synthetic data for fitness
VisionFit API
Computer vision is revolutionizing the world of remote fitness and remote physical therapy by bringing AI-powered trainers to our homes.
Applications
2D and 3D pose estimation
Rep counting
Form correction
Person segmentation
Activity classification
Dumbbell detection
Empowering these use cases
Never be blocked by lack of data. Build computer vision models for these use cases (and more!).
Rep counting
Automatically count reps across a variety of exercises by training against perfect per-frame rep labels in the synthetic data.
Form correction
Detect when users are performing exercises with good or bad form and provide coaching tips as needed.
2D and 3D pose estimation
Train accurate pose models for an arbitrary number of keypoints using synthetic data.
Person segmentation
Automatically segment the person from their environment for interesting real-time visualizations.
Activity detection
Detect if someone is taking a break or actively doing a specific exercise.
Dumbbell detection
Automatically detect if dumbbells are being used and what their weights are.
Features
Build better models faster with feature-rich synthetic data from the Fitness APIs.
Lifelike rep behavior
Each rep is done slightly differently. Just like a real human. Seed movements are based on real-world movement data in order to model the unique intricacies of movement and produce realistic and robust training data.
Furthermore, the API gives users the ability to add vast amounts of movement variation. From cadence variation to movement truncations, no two reps are done exactly the same way.
Any lighting condition
Pick from a library of realistic 3D indoor environments or import your own. For each scene, control the lighting conditions and camera position so that videos have varied and realistic shadows. Indoor and outdoor lighting conditions can be controlled separately.
Occlusion and clutter
Real-world remote fitness scenes are typically people in small spaces with lots of furniture and objects around them. Mimic realistic occlusions in the synthetic data via the API occlusion parameter (0-60% of the avatar).
Let the API select random objects, pick specific ones from the Infinity library, or import your own.
Toggle an API parameter to turn a scene from neat to messy by adding household clutter. Stress test your models since clutter typically makes CV problems more challenging.
Digital wardrobe
Generate synthetic training data of avatars with diverse body shapes, skin tones, and clothing. Pick skin-tight or baggy clothing from the Infinity wardrobe. Control the amount of bagginess in an item of clothing using
the API.
Synthetic videos will have segmentation and bounding box labels around clothing items, in addition to perfect keypoint annotations for the joint positions underneath.
Import any exercise
Import a specific exercise into the API by using the Infinity Auto-import tool. Simply upload a regular RGB video and extract its 3D motion. Once in the API, the motion trajectory can be applied to any avatar in any location. Kinematic variation is automatically added by the Infinity platform. Learn more
Resources
Make synthetic data your competitive edge
Have questions? Want to book a scoping session?
Reach out to info@infinity.ai.