Knowledge driven weights estimation for large-scale few-shot image recognition.

•Propose a knowledge-based large-scale few-shot learning framework that leverages relations between larges numbers of many-shot classes and few-shot classes to estimate the initialization parameters of classifiers for few-shot classes.•On the knowledge graph, both semantic relations (WordNet hierarchy) and visual relations (visual similarities) are explored for few-shot learning. As far as we know, this is the first attempt to directly encode the visual relations in the knowledge graph for few-shot classification.•Verify the effectiveness of the proposed framework on different splits of un-seen addtional ImageNet dataset.