Prototype rectification for zero-shot learning

Yuanyuan Yi, Guolei Zeng, Bocheng Ren, Laurence T. Yang, Bin Chai, Yuxin Li

Abstract: Highlights•Semantic-visual alignment can alleviate the domain shift in zero-shot learning.•Meta-learning can enhance the generalizability of zero-shot learning models.•The attention mechanism still works in zero-shot learning.•The alignment between semantic and visual modalities requires mutual guidance.•The naive embedding-based method can achieve excellent zero-shot recognition.