A robust transductive distribution calibration method for few-shot learning

Few-shot learning (FSL) has gained much attention and has recently made substantial progress. To alleviate the data constraints in FSL, previous studies have attempted to generate features by learning a feature distribution. However, the learned distribution is biased and unstable due to limited labeled data, and the features from it can be even more biased, which decreases its generalizability. This paper proposes a Robust Transductive Distribution Calibration (RTDC) method to estimate feature distributions of few-shot classes in a more accurate and robust way. First, we capture the underlying distribution information by precisely estimating the covariance matrix of each novel category. Second, we consider the distribution similarity between labeled and unlabeled samples using the estimated covariance matrix and then optimize the feature distribution in a transductive manner. Extensive experiments demonstrated the effectiveness and significance of our method on several FSL benchmarks, including miniImageNet, tieredImageNet, CUB, and CIFAR-FS.