This database contains automatically extracted features (head pose, gaze, speaking status, people/object positions) as well as manual annotations (visual focus of attention, grasp/release actions) from three different datasets:
The conversation datasets were manually annotated with the visual focus of attention (around 3000 annotated frames per video in the UBImpressed dataset and 9000 annotated frames per video in KTH-Idiap Group-Interviewing dataset. The ManiGaze dataset provides VFOA ground truth for its gazing sessions (51 target per subject) and actions annotations for its object manipulation session (22 actions per subject). See the reference paper for details on feature extraction.
This database was mainly collected to make experiments on visual focus of attention estimation and gaze estimation calibration.
Rémy Siegfried and Jean-Marc odobez, « Robust Unsupervised Gaze Calibration using Conversation and Manipulation Attention Priors », ACM Transactions on Multimedia Computing, Communications, and Applications, 2021
https://publications.idiap.ch/index.php/publications/show/4611
https://publications.idiap.ch/index.php/publications/show/4294