Abstract
Despite the span in estimating pain from facial expressions, limited workshave focused on estimating the sequence-level pain, which is reported bypatients and used commonly in clinics. In this paper, we introduce a novelStatistical Relation Network, referred to as PainNet, designed for theestimation of the sequence-level pain. PainNet employs two key modules, theembedding and the relation modules, for comparing pairs of pain videos, andproducing relation scores indicating if each pair belongs to the same paincategory or not. At the core of the embedding module is a statistical layermounted on the top of a RNN for extracting compact video-level features. Thestatistical layer is implemented as part of the deep architecture. Doing so,allows combining multiple training stages used in previous research, into asingle end-to-end training stage. PainNet is trained using the episode-basedtraining scheme, which involves comparing a query video with a set of videosrepresenting the different pain categories. Experimental results show thebenefit of using the statistical layer and the episode-based training in theproposed model. Furthermore, PainNet outperforms the state-of-the-art resultson self-reported pain estimation.