Reproduce “From Facial Expression Recognition to Interpersonal Relation Prediction” on OpenResearch

About this paper

Zhanpeng Zhang, Ping Luo, Chen Change Loy, Xiaoou Tang

This research investigates the characterization and quantification of fine-grained interpersonal relation traits directly from face images in uncontrolled settings. It develops a two-stage deep learning framework that includes a novel attribute propagation method for robust learning from heterogeneous datasets and a Siamese network for pairwise relation prediction, achieving improved performance on facial attribute and expression recognition, and successfully inferring complex social dynamics.

Key insights

• Fine-grained interpersonal relation traits, based on psychological models like Kiesler's interpersonal circle, can be robustly inferred from face images captured in unconstrained environments.
• A novel attribute propagation method effectively leverages heterogeneous and incompletely labeled datasets, improving the generalization and robustness of learned facial representations.
• Pre-training deep convolutional networks on facial attributes, particularly expressions, and incorporating spatial cues between face pairs are crucial for accurate interpersonal relation prediction.

Summary by alphaXiv.