PyData London 2023

Object detection is arguably the most common Computer Vision task. It is applied to images and videos across various domains. However, action recognition is a tad different from object detection because it can be difficult to tell certain actions from a single image. It is hard to tell if a door is being opened or closed or tell what martial art technique is being executed from an image.

In this tutorial, the MMAction2 framework will be used to train an action recognition model to detect what Judo throws are being performed in videos. While it will be fun seeing Machine Learning techniques applied to Martial Arts, the knowledge and techniques applied can easily be generalized to other action recognition tasks where simple object detection does not suffice.

The target audience should have some experience working on object detection tasks, as this offers a lot more context as to why action recognition is a tricky problem. However, this knowledge is not compulsory to participate in the session.

The session will be set up to provide access to already labelled videos for the audience, then they can walk through the provided notebook to train their own Judo action recognition models during the session.

At the end of this tutorial, the audience will:

1. Understand the difference between object detection and action recognition tasks
2. Learn how to use the MMAction2 library
3. Be able to train a model to recognize Judo throws

In order to showcase the adaptability of the techniques to other kinds of videos, there will be a separate notebook for recognizing other actions so the audience can see how easy it is to modify the models. While they will know more about action recognition from this presentation, as a takeaway, they will get some look into the world of Judo.