Do they Share the Same Tail? Learning Individual Compositional Attribute Prototype for Generalized Zero-Shot Learning

Yuyan Shi, Chenyi Jiang, Run Shi, Haofeng Zhang; Proceedings of the Asian Conference on Computer Vision (ACCV), 2024, pp. 4595-4611

Attributes are considered fundamental in zero-shot learning. By incorporating the correspondences between classes and attributes as prior knowledge, the model is able to approximate a class prototype for numerous classes without the need for any visual samples of these classes. In the majority of prior research, attributes are considered primitives and are not subjected to further subdivision. While the only distinction between shared attributes across classes is the absolute magnitude of their values, this does not adequately reflect the more significant visual differences between these classes in natural images. To address this issue, we propose learning the Individual Compositional Attribute Prototype (InCAP). Specifically, InCAP does not treat attributes as the sole primitives but uses attribute semantics as objects in compositions, while class semantics are introduced as a special kind of state description within these compositions. This approach allows attributes and classes to form the structure of the composition. To avoid domain shift when encountering unseen classes, these compositional attributes are not used for direct contrusting class prototypes. Instead, they serve as spatial composition bottlenecks to suppress potential overfitting caused by attribute-visual mismatches during training and provide advanced location guidance information during testing. Experiments demonstrate that InCAP achieves leading results on mainstream datasets, validating the full potential of this strategy.