RHPENet: Robust Head Pose Estimation by Learning Heterogeneous Relationships from Facial Region Cues in Human Robot Interaction

Abstract

<p>Head pose estimation (HPE) techniques frequently encounter difficulties when handling extreme angles, occlusions, and uneven lighting conditions. In this paper, we present a novel heterogeneous relationship learning framework designed to mitigate these limitations by exploiting facial regions of interest (FRoI) and their complex interdependencies. Our approach stems from two fundamental discoveries: first, the critical importance of FRoI for pose determination, and second, the heterogeneous relationship between neighboring postures. The proposed architecture consists of three main modules: region feature generator (RFG), hierarchical structural representation (HSR), and cross-relation aggregator (CRA). The RFG incorporates an adaptive attention mechanism that prioritizes diagnostically significant facial zones. Within the HSR component, we implement a novel "Rugby-style" cross-level connectivity pattern to enhance feature integration. The CRA employs Transformer-based techniques to uncover both spatial and angular dependencies. Comprehensive evaluations conducted on major HPE benchmarks (300W_LP, AFLW2000, and BIWI) demonstrate that our RHPENet model consistently outperforms existing approaches.<br></p>