Hack Pompey: Move ideas

Using computer vision, can you build a tool that is able to recognise the elevation of the ball in a screenshot or live stream of a football match?

Getting Started

It’s possible to train a computer vision model to "see" features of an image to then classify or bucket them as appropriate. This kind of problem is called "Image Classification".

In this challenge, you will be training a computer vision model to distinguish whether the football in a match screenshot is on the ground, in mid-elevation, or at a high-elevation.

We’ve provided the labelled data that you will need to train your model. The data has already been split into "train", "validation" and "test" sets. In the labelled data, the 1st number is the class, and the other 4 are bounding-box co-ordinates of the ball. The classes are: 0 - ball on the ground, 1 - ball is mid-elevation, 2 - ball is high-elevation.

You can download the labelled training data from this Google Drive page (It’s a 500Mb zip file).

Tips

• TensorFlow Hub and Keras Applications have pre-trained computer vision models. These models have already been taught to “see”, some could be fine-tuned for the task or you could use their outputs (observations or image features) as inputs for your own logic or model.
• OpenCV and Keras have powerful tools for image processing and data augmentation.
• Pandas and sci-kit-learn are great for general data handling and evaluation. (hint: you might want ski-kit-learn classification_report )
• Google Colab lets you run Python scripts in the cloud using a GPU for free. Many tutorials and libraries provide Jupyter notebooks you can open, run, and customise in Colab.

If you’re stuck on where to start or need help getting over any hurdles, you can tag our resident AI Expert Sage Ralph (@growingsage) on our Discord, or speak to one of the Hack Pompey Staff and we’ll be happy to help!