Answers are grounded in this paper and cite their sources.
Few-shot learning aims to recognize new-class items under the circumstances with a few labeled support samples. However, many methods may suffer from poor guidance of limited new-class samples that are not suitable for being regarded as class centers. Recent works use word embedding to enrich the new-class distribution message but only use simple mapping between visual and semantic features during training. To solve the aforementioned problems, we propose a fusion method that constructs a class relation graph by semantic meaning as guidance for feature extraction and fusion, to help the learning of the second-order relation information, with a light training request. In addition, we introduce two ways to generate pseudo prototypes for augmentation to resolve the lack of representation due to limited samples in novel classes: 1) A Generation Module(GM) that trains a small structure to generate visual features by using word embedding; 2) A Relation Module(RM) for training-free scenario that uses class relations in semantics to generate visual features. Extensive experiments on benchmarks including miniImageNet, CIFAR-FS and FC-100 prove that our fusion method achieves state-of-the-art results.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: