Interpersonal relation defines the association, e.g., warm, friendliness, anddominance, between two or more people. Motivated by psychological studies, weinvestigate if such fine-grained and high-level relation traits can becharacterized and quantified from face images in the wild. We address thischallenging problem by first studying a deep network architecture for robustrecognition of facial expressions. Unlike existing models that typically learnfrom facial expression labels alone, we devise an effective multitask networkthat is capable of learning from rich auxiliary attributes such as gender, age,and head pose, beyond just facial expression data. While conventionalsupervised training requires datasets with complete labels (e.g., all samplesmust be labeled with gender, age, and expression), we show that thisrequirement can be relaxed via a novel attribute propagation method. Theapproach further allows us to leverage the inherent correspondences betweenheterogeneous attribute sources despite the disparate distributions ofdifferent datasets. With the network we demonstrate state-of-the-art results onexisting facial expression recognition benchmarks. To predict inter-personalrelation, we use the expression recognition network as branches for a Siamesemodel. Extensive experiments show that our model is capable of mining mutualcontext of faces for accurate fine-grained interpersonal prediction.
展开▼
