Learning object-centric complementary features for zero-shot learning

Highlights：

• An object-centric complementary feature (OCF) learning model is proposed for zero-shot learning.

• The object-centric region refine network can automatically discover the object-centric region and obtain fine-scale images without any human annotation.

• The GDA module is proposed to capture more global visual features corresponding to semantic knowledge like ‘four legs’ and ‘two wings’.

• The proposed LDA module can discover the subtle visual difference between similar categories and contributes to better performance on fine-grained dataset.

• This work provides an end-to-end trainable novel model to capture more discriminative visual features with the guidance of semantic knowledge and compatibility loss.