Experiments in cross-domain few-shot learning for image classification

摘要

Cross-domain few-shot learning has many practical applications. This paper attempts to shed light on suitable configurations of feature exactors and 'shallow' classifiers in this machine learning setting. We apply ResNet-based feature extractors pretrained on two versions of the ImageNet dataset to five target domains with different degrees of similarity to ImageNet, varying the feature extractor size, the network stage at which features are extracted, and the learning algorithm applied to the extracted features. We evaluate standard teaming algorithms such as logistic regression and linear discriminant analysis, as well as variants thereof, and additionally consider the effect of normalising the feature vectors using various p-norms. We also apply multi-instance learning to improve training image utilisation. In our experiments, the cosine similarity classifier and l(2) -regularised 1-vs-rest logistic regression generally exhibit the best classification performance. We also find that algorithms such as linear discriminant analysis yield consistently higher accuracy using l(2)-normalised feature vectors. Features extracted from the penultimate stage of a ResNet-101 model, and multi-instance learning techniques, produce the highest accuracy for most target domains. Our results will inform practitioners who are considering the application of pretrained ImageNet feature extractors in cross-domain few-shot settings.