Zero-Shot Classification at Different Levels of Granularity
Zero-shot classification (ZSC) is the task of learning predictors for classes not seen during training. The different methods proposed in literature are evaluated over specific datasets with their specific class partitions, but little attention has been paid to the impact of the dataset granularity when ZSC is performed. The novelty of this work is to generate synthetic datasets by controlling their granularity level to analyze the ZSC performance afterwards. Moreover, it presents an approach that allows us to preserve the visual and semantic structures. The experiments show that ZSC performance exhibits strong differences depending on the data granularity and it reveals the relevance of both visual and semantic spaces when performing ZSC.