Label Hallucination for Few-Shot Classification

Few-shot classification requires adapting knowledge learned from a large annotated base dataset to recognize novel unseen classes, each represented by few labeled examples. In such scenario, pretraining network with high capacity on the and then finetuning it examples causes severe overfitting. At same time, training simple linear classifier top of ``frozen'' features fails adapt model properties effectively inducing underfitting. this paper we propose an alternative approach both these two popular strategies. First, our method pseudo-labels entire using trained classes. This ``hallucinates'' classes in dataset, despite categories not being present database (novel are disjoint). Then, finetunes distillation loss pseudo-labeled examples, addition standard cross-entropy dataset. step trains contextual appearance cues that useful for novel-category recognition but large-scale thus overcoming inherent data-scarcity problem few-shot learning. Despite simplicity approach, show outperforms state-of-the-art four well-established benchmarks. The code is available at https://github.com/yiren-jian/LabelHalluc.