SpatialFormer: Semantic and Target Aware Attentions for Few-Shot Learning

Recent Few-Shot Learning (FSL) methods put emphasis on generating a discriminative embedding features to precisely measure the similarity between support and query sets. Current CNN-based cross-attention approaches generate representations via enhancing mutually semantic similar regions of pairs. However, it suffers from two problems: CNN structure produces inaccurate attention map based local features, backgrounds cause distraction. To alleviate these problems, we design novel SpatialFormer more accurate global features. Different traditional Transformer modeling intrinsic instance-level which causes accuracy degradation in FSL, our explores semantic-level pair inputs boost performance. Then derive specific modules, named Semantic Attention (SFSA) Target (SFTA), enhance target object while reduce background Particularly, SFSA highlights with same information SFTA finds potential foreground feature that are base categories. Extensive experiments show effective achieve new state-of-the-art results few-shot classification benchmarks.