Decomposed Prototype Learning for Few-Shot Scene Graph Generation

• DOI: 10.1145/3700877
• Publication Date: 2024-10-21T15:55:21Z
• Authors: Xingchen Li; Jun Xiao; Guikun Chen; Yinfu Feng; Yi Yang; An-An Liu; Long Chen
• Keywords: FOS: Computer and information sciences; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition

Today's scene graph generation (SGG) models typically require abundant manual annotations to learn new predicate types. Therefore, it is difficult apply them real-world applications with massive uncommon categories whose are hard collect. In this paper, we focus on Few-Shot SGG (FSSGG) , which encourages be able quickly transfer previous knowledge and recognize unseen predicates well only a few examples. However, current methods for FSSGG hindered by the high intra-class variance of in SGG: On one hand, each category commonly has multiple semantic meanings under different contexts. other visual appearance relation triplets same differs greatly subject-object compositions. Such great inputs makes generalizable representation few-shot learning (FSL) methods. found that highly related composed subjects objects. To model context, propose novel Decomposed Prototype Learning (DPL) FSSGG. Specifically, first construct decomposable prototype space capture diverse semantics patterns objects decomposing into prototypes. Afterwards, integrate these prototypes weights generate query-adaptive more reliable query sample. We conduct extensive experiments compare various baseline show effectiveness our method.