Dataset Bias in Few-shot Image Recognition
share
The goal of few-shot image recognition (FSIR) is to identify novel categories with a small number of annotated samples by exploiting transferable knowledge from training data (base categories). Most current studies assume that the transferable knowledge can be well used to identify novel categories. However, such transferable capability may be impacted by the dataset bias, and this problem has rarely been investigated before. Besides, most of few-shot learning methods are biased to different datasets, which is also an important issue that needs to be investigated in depth. In this paper, we first investigate the impact of transferable capabilities learned from base categories. Specifically, we introduce relevance to describe relations of base and novel categories, along with instance diversity and category diversity to depict distributions of base categories. The FSIR model learns better transferable knowledge from relative training data. In the relative data, diverse instances or categories can further enrich the learned knowledge. We conduct experiments on different sub-datasets of ImagNet, and experimental results demonstrate category relevance and category/instance diversity can depict transferable bias from distributions of base categories. Second, we investigate performance differences on different datasets from dataset structures and few-shot learning methods. Specifically, we introduce image complexity, inner-concept visual consistency, and inter-concept visual similarity to quantify characteristics of dataset structures. We use these quantitative characteristics and four few-shot learning methods to analyze performance differences on five different datasets. Based on experimental analysis, some insightful observations are obtained from the perspective of both dataset structures and few-shot learning methods. These observations are useful to guide future FSIR research.
READ FULL TEXT
