Deformable Locality-Coordination Graph Motifs for 3D Skeleton Based Person Re-Identification

Existing 3D skeleton based person re-identification (re-ID) approaches typically model skeletons as graphs to capture body relations and motion. However, they often rely on <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">fixed</i> joint’s connections such as adjacency for relation modeling, while lacking a flexible and specific focus on key body joints or parts of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">different levels</i> to capture various local relations (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">“locality”</i>) and limb relations (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">“coordination”</i>). In this letter, we propose Deformable Locality-Coordination graph Motifs (DL-CM) that can guide the body relation learning to particularly capture multi-order <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">locality</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">coordination</i> of key gait-specific body parts to enhance person re-ID performance. Specifically, we first devise Deformable Locality Motifs (DLM), which are applicable to deformed skeleton graphs at different levels, to simultaneously focus on different-order neighbors’ relations for body structure and pattern learning. Then, we propose Deformable Coordination Motifs (DCM) to concurrently capture local and global coordination of different-level limbs in deformed graphs, so as to facilitate learning discriminative gait patterns for person re-ID. Extensive experiments on four public benchmarks demonstrate the effectiveness of DL-CM on state-of-the-art models and different-level graph representations to improve person re-ID performance.