The evaluation server is now open for the next speaker prediction task
We are happy to announce that the test server for the next speaker prediction task can be used for evaluation now. If your method uses the audio modality for next speaker prediction, please be aware that the volume of the audio can vary between recordings. This might be relevant if you rely on features that encode the volume.
Furthermore we would like to inform you that baseline results for the eye contact detection task are in the leaderboard. For the eye contact baseline, we extracted head pose and gaze features wiht OpenFace 2.0 and trained a separate RBF-SVM for each seating position that predicts the eye contact classes. This baseline reached 0.52 Accuracy on the test set.
If you have any troubles with the evaluation system (building images, submitting them, interpreting the output,...), please do not hesitate to contact us for help!