Teaching a model to read the game

A SaaS B2B business working with Cloud Combinator on AWS. The client is anonymised at their request.

The challenge had three focus areas, each one a step from raw images towards automated understanding.

Curated, labelled the client's images are used to train an Amazon Rekognition Custom Labels model that learns to recognise the categories of play that matter to the client trained, the model is deployed and ready to classify images on demand.

The engagement delivered a working, automated image classification pipeline. A custom vision model was trained on curated the client's imagery and deployed behind a serverless workflow, so that every new image is classified and its prediction stored without manual effort. Evaluating Rekognition Custom Labels alongside a SageMaker classifier meant the final approach was chosen on evidence.

• Clean dataset prepared — The client's imagery was curated and labelled, with weaker images excluded, to give the model clear examples to learn from.
• Custom model trained — An Amazon Rekognition Custom Labels model was trained to recognise the types of play in an image and returns a label with a confidence score.
• Automated pipeline delivered — A S3 upload triggers a Lambda function that classifies the image and stores the prediction, with no manual step.
• Handover — The client left the engagement with a deployed model, a working inference pipeline and a clear, evidence-based view of the Rekognition and SageMaker options.

With a working classifier in place, the natural next steps (projected) would be to extend the label set to cover more phases of play, grow the training data to sharpen accuracy, and feed the classified output into the client's wider analytics and content tooling.