Comprehensive Analysis of Few-shot Image Classification Method Using Triplet Loss

Image classification task is a very  important problem of a computer vision area. The first approaches to image classification tasks belong to a classic straightforward algorithm. Despite the successful applications of such algorithms a lot of image classification tasks had not been solved until machine learning approaches were involved in a computer vision area. An early successful result of machine learning applications helps researchers with extracted features classification which was not available without machine learning models. But handcrafter features were required which left the most complicated classification task impossible to solve. Recent success in deep learning allows researchers to implement automatic trainable feature extraction. This gave significant progress in the computer vision area last but not least. Processing large-scale datasets bring researchers great progress in automatic feature extraction thus combining such features with precious approaches led to groundbreaking in computer vision. But a new limitation has come - dependency on large amounts of data. Deep learning approaches to image classification task usually requires large-scale datasets. Moreover, modern models lead to unexpected behavior in distribution datasets. A few-shot learning approach of deep learning models allows us to dramatically reduce the amount of required data while keeping the same promising results. Despite reduced datasets, there is still a tradeoff between the amount of available data and trained model performance. In this paper, we implemented a siamese network based on triplet loss. Then, we investigate a relationship between the amount of available data and few-shot model performances. We compare the models obtained by metric-learning with baselines models trained using large-scale datasets.