Fine-grained few shot learning with foreground object transformation

Traditional fine-grained image classification generally requires abundant labeled samples to deal with the low inter-class variance but high intra-class problem. However, in many scenarios we may have limited for some novel sub-categories, leading few shot learning (FG-FSL) setting. To address this challenging task, propose a method named foreground object transformation (FOT), which is composed of extractor and posture generator. The former aims remove background, tends increase difficulty as it amplifies while reduces variance. latter transforms generate additional sub-category. As data augmentation method, FOT can be conveniently applied any existing algorithm greatly improve its performance on FG-FSL tasks. In particular, combination FOT, simple fine-tuning baseline methods competitive state-of-the-art both inductive setting transductive Moreover, further boost performances latest excellent bring them up new state-of-the-art. addition, also show effectiveness general FSL