A comprehensive survey of zero-shot image classification: methods, implementation, and fair evaluation

Deep learning methods may decline in their performance when the number of labelled training samples is limited, in a scenario known as few-shot learning. The methods may even degrade the accuracy in classifying instances of classes that have not been seen previously, called zero-shot learning. While the classification results achieved by the zero-shot learning methods are steadily improved, different problem settings, and diverse experimental setups have emerged. It becomes difficult to measure fairly the effectiveness of each proposed method, thus hindering further research in the field. In this article, a comprehensive survey is given on the methodology, implementation, and fair evaluations for practical and applied computing facets on the recent progress of zero-shot learning.