Distilling interaction knowledge for semi-supervised egocentric action recognition

Egocentric action recognition, the identification of actions within video content obtained from a first-person perspective, is receiving increasing attention due to the widespread adoption of wearable camera technology. Nonetheless, the task of annotating actions within a video characterized by a cluttered background and the presence of various objects is labor-intensive. In this paper, we consider learning for egocentric action recognition in a semi-supervised manner. Inspired by the fact that videos captured from first-person viewpoint usually contain rich contents about how human hands interact with objects, we thus propose to employ a popular teacher-student framework and distill the interaction knowledge between hand and objects for semi-supervised egocentric action recognition. We refer to the proposed method as Interaction Knowledge Distillation or IKD. Specifically, the teacher network takes hands and action-related objects in the labeled videos as input, and uses graph neural networks to capture their spatial-temporal relations as graph edge features. The student network then takes the detected hands/objects from both labeled and unlabeled videos as input and mimics the teacher network to learn from the interactions to improve model performance. Experiments are performed on two popular egocentric action recognition datasets, Something-Something-V2 and EPIC- KITCHENS-100, which show that our proposed approach consistently outperforms recent state-of-the-art methods in typical semi-supervised settings.