Habeeb Shopeju - Martial Arts Meets Machine Learning: Recognizing Judo Throws with MMAction2

Learn how machine learning and action recognition come together in a innovative project recognizing judo throws with MMAction2, using the `mm-action` library, `TSN` model, and data augmentation techniques.

Key takeaways

The project aims to recognize judo throws using machine learning and action recognition.
The data set is in ANM file format, which contains a text file that defines the actions.
The mm-action library is used to develop the model, which supports various image and video formats.
The TSN model is used, which is a 3D convolutional network that processes video data in both spatial and temporal domains.
Data augmentation is used to increase the diversity of the training data, including random cropping, flipping, and color jittering.
The model is trained using a custom config file, which defines the hyperparameters and training parameters.
The model is evaluated using a validation set, and the best-performing model is selected based on the validation loss.
The model is then used to perform action recognition on new, unseen videos.
The project uses the OpenPose library to extract pose information from the video frames.
The project also uses the pytorch library to develop the model and perform training and inference.
The project is a personal project developed by the author, who is a research engineer at the Thompson Reuters Lab.

Habeeb Shopeju - Martial Arts Meets Machine Learning: Recognizing Judo Throws with MMAction2

More talks