Few-Shot Classification with Semantic Augmented Activators

Metric-based methods predict class labels by measuring the distance between a few given samples, often failing to preserve more useful semantic details in their vectorial representations. In this paper, we propose Semantic Augmented Activators (SAA), which are generated based on the variance of the intra-set samples in an unsupervised manner, to enhance the discriminability of feature vectors with more class-related semantic information. This generation process does not rely on any learnable parameters. Meanwhile, to align the SAA preferred to operate in the intra-set and sufficiently leverage the finite samples, we treat the Self-Cross loss as an auxiliary loss, which bi-directionally complements the limitations of the traditional loss function. Additionally, we introduce Map-To-Cluster, a transductive module to map the SAA-enhanced features to a lower-dimensional embedding space. This encourages proximity among similar samples and separation among dissimilar samples. The resulting methods are lightweight and computationally efficient. Our methods demonstrate competitive performance on the mini-ImageNet and tiered-ImageNet benchmarks, and achieve outstanding results in Cross-Domain Few-Shot classification.