Multimodal Joint Head Orientation Estimation in Interacting Groups via Proxemics and Interaction Dynamics

Human head orientation estimation has been of interest because serves as a cue to directed social attention. Most existing approaches rely on visual and high-fidelity sensor inputs deep learning strategies that do not consider the context unstructured crowded mingling scenarios. We show alternative inputs, like speaking status, body location, orientation, acceleration contribute towards estimation. These are especially useful in in-the-wild settings where features either uninformative due occlusions or prohibitive acquire physical space limitations concerns ecological validity. argue such needs account for physically evolving interaction formed by all individuals group. To this end, we propose an LSTM-based method combines hidden representations group members. Our framework jointly predicts orientations members is applicable groups different sizes. explain contribution modalities model performance The proposed outperforms baseline methods explicitly context, generalizes unseen dataset from event.