Hands-on workshop on gaze and interaction detection on videos using deep learning

Dear friends, we are happy to announce a workshop that will take place on the 20-21st of Nov 2025, 9:00-16:00 at Raum E0.14, DPZ.

This two-day workshop offers a practical introduction to training deep learning models for gaze and interaction detection.

On Day 1, participants will learn how to select suitable video frames, create and label training datasets, train and apply a tracking model.
On Day 2, we will use these tracking outputs to annotate social interactions and train a gaze or interaction detection model.

The workshop is designed for SFB research groups interested in applying automated detection of interactions (e.g. gaze) to their own videos. If you do not have a video dataset (yet) but would still like to try out the models, we can provide datasets. No previous coding knowledge is required to participate in the workshop.

Interactions can happen between two individuals or between an individual and an object in the scene. You as a human should be able to reliably detect the interaction. Current models only allow training one interaction at a time (e.g. "looking-at" or "interacting-with" or "grooming"), and as the models are trained on single images, we expect low success rates for interactions that highly depend on temporal context. The interactions should not be extremely rare; at least 50 different occurrences of the interaction should be found in your videos.