Dataset Bias in Few-Shot Image Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot image recognition (FSIR) aims to identify novel categories using limited data by leveraging knowledge from base categories.
• Current FSIR methods often assume transferable knowledge is readily usable, overlooking the impact of dataset bias.
• Dataset bias and method-specific biases are critical, yet under-investigated, issues affecting FSIR performance.

Purpose of the Study:

• To investigate the influence of dataset bias on the transferability of knowledge in FSIR.
• To analyze performance variations across different datasets and few-shot learning methods.
• To provide insights for guiding future FSIR research.

Main Methods:

• Quantified relationships between base and novel categories using category relevance.
• Analyzed base category distributions via instance density and category diversity.
• Introduced image complexity, intra-concept visual consistency, and inter-concept visual similarity to characterize dataset structures.
• Evaluated eight few-shot learning methods across multiple datasets.

Main Results:

• Category relevance, instance density, and category diversity effectively depict transferable bias from base category distributions.
• Dataset structure characteristics and specific few-shot learning methods explain performance differences.
• Experimental results on ImageNet sub-datasets validate the proposed analysis metrics.

Conclusions:

• Dataset characteristics significantly influence the effectiveness of transferable knowledge in FSIR.
• Understanding dataset bias is crucial for developing robust and generalizable few-shot learning models.
• The findings offer guidance for selecting appropriate methods and datasets in future FSIR research.