Collegiate athletics coaches spend significant time manually reviewing practice footage (approximately 10 hours weekly for the University of Pittsburgh diving team), which limits their ability to focus on direct athlete development and creates a critical need for automation in sports video processing.
Components Built:
- Zero-shot vision-language model (VLM) for automated video processing
- Automated detection system for identifying individual dives from continuous footage
- Highlight reel generation system that processes 2.5-hour practice sessions
- Open-source framework designed to handle real-world constraints (single-angle footage, background activity, water reflections)
The solution was powered by several key AWS services:
• Amazon S3 – Primary storage for video uploads (videos/ prefix) and processed results (results/ prefix)
• AWS Lambda – Orchestration service that triggers when videos are uploaded to S3 and launches the processing pipeline
• Amazon ECS – Container orchestration service that runs the video processing tasks using Docker containers
• Amazon EC2 – Provides GPU-enabled instances (specifically g4dn.2xlarge with NVIDIA T4 GPU) for the compute-intensive video analysis
• Amazon ECR – Container registry for storing the Docker images containing the video processing application
• AWS CDK – Infrastructure as Code framework used for deploying and managing all the AWS resources
The architecture follows an event-driven pattern where S3 upload events trigger Lambda functions, which then launch containerized processing jobs on GPU-enabled EC2 instances through ECS, making it a scalable and cost-effective solution for automated video processing.
Visit the blog post to learn more.
Access to the GitHub repository.