Hello Google AI Developers!!
My name is Joel Ownby, and on behalf of the Willow Sports Intelligence team, I’m excited to share our project with the Google Developer Community. We’ve built a platform to transform sports videos into personalized, narrative training experiences that learn, adapt, and grow.
The Problem We’re Solving
The youth sports market faces a critical shortage of qualified coaches, exacerbated by economic and geographic divides, where over 30 million children in the US lack access to private training, and more affluent and engaged families spend $800 - $2500 annually per child on specialized instruction.
What is Willow?
Willow is a “GenAI coach that feels remarkably human.” Users can upload videos of their performance, and our AI provides a deep analysis of their mechanics, play, and strategy. This isn’t just about tracking; it’s about interpretation, reasoning, and adaptation provided in a narrative human-like coaching experience.
Here’s a short video showcasing some kinematic playback: https://www.youtube.com/shorts/txewWekZvzM
Here’s a longer video showcasing a golf analysis: https://youtu.be/3rHmXItMFjQ?si=1PKrM91-TTmwpgAz
Currently, we have two primary products:
- GolfLogic: Provides an unparalleled analysis of a golfer’s swing mechanics, play, and strategy directly from video.
- BigHitter: Delivers a comprehensive analysis of baseball and softball mechanics for hitting, pitching, fielding, and catching.
Our AI Capabilities
Our platform is built on a foundation of advanced computer vision and AI technologies to understand the game. Here are some of the core capabilities we’ve developed, which are crucial for sports analysis:
-
Object Detection & Tracking: We identify what’s on the field and track the movement of players and the ball over time. We use Multi-Object Tracking (MOT) to follow numerous players at once, even in dense, fast-paced scenarios.
-
Pose Estimation & Action Recognition: We use ‘skeletal tracking’ to gain invaluable insights into an athlete’s biomechanics, joint angles, and technique.
-
Event Detection & Video Summarization: Our system automatically identifies significant game occurrences (like a pitch release or ball contact) to streamline highlight generation.
-
Data Overlay: We can overlay key insights—like tracking paths or performance data—directly onto the video to make the analysis clear and actionable.
Our Tech Stack - The “How”
This is a developer forum, so we’d love to talk about how it’s built. Our recent release notes mention significant enhancements to our core system architecture, and we’re excited to discuss it.
This is where we’d love to get into a discussion with the community. We are constantly thinking about the best tools for the job.
-
AI & Computer Vision: Our models, frameworks, application layer, architecture, memory structures, etc. are the core of our IP. We are always interested in hearing how others in the community are managing their own ML pipelines and model deployment, particularly for video processing.
-
Cloud Infrastructure: Powering this kind of analysis requires a robust and scalable backend. We’d be happy to answer questions about the infrastructure choices we made to handle large video uploads and intensive computational workloads.
-
Front End: We’re focused on delivering a clean, intuitive user experience.
Questions for the Community:
We’re here to share and learn.
-
What interesting or challenging technical problems you see in the sports technology space or in our model specifically?
-
For those working with video analysis, how have you approached the trade-off between analysis depth, processing speed, and cost?
-
Any feedback or questions on the features we’ve presented?
Thanks for checking out Willow Sports Intelligence!
Links to patch notes & website in first comment.