Learning a Latent Space with Triplet Network for Few-Shot Image Classification

Few-shot image classification has attracted much attention due to its requirement of limited training data for target classes. Existing methods usually pretrain a network with images from the base set as feature extractor to obtain features of images from novel set. However, the pretrained feature extractor cannot extract accurate representation for categories have never seen, making images from novel set difficult to distinguish. To be specific, in the pretrained feature space, there exist a large number of overlapped areas between novel categories. To address this issue, it is crucial to acquire a space, where features from same class are gathering together and features from different classes are far away from each other. Since lots of experiments have proved that the triplet network is effective to achieve this goal, in this paper, we base our network on the Maximum a posteriori (MAP), learning a latent space with triplet network to project features from pretrained feature space into a more discriminative one. Experimental results on four few-shot benchmarks show that it significantly outperforms the baseline methods, improves around 1.09%∼13.09% than the best results in each dataset on both 1- and 5-shot tasks.