- TL;DR: We address the problem of unsupervised few-shot object recognition, where all training images are unlabeled and do not share classes with test images.
- Abstract: This paper addresses unsupervised few-shot object recognition, where all training images are unlabeled and do not share classes with labeled support images for few-shot recognition in testing. We use a new GAN-like deep architecture aimed at unsupervised learning of an image representation which will encode latent object parts and thus generalize well to unseen classes in our few-shot recognition task. Our unsupervised training integrates adversarial, self-supervision, and deep metric learning. We make two contributions. First, we extend the vanilla GAN with reconstruction loss to enforce the discriminator capture the most relevant characteristics of "fake" images generated from randomly sampled codes. Second, we compile a training set of triplet image examples for estimating the triplet loss in metric learning by using an image masking procedure suitably designed to identify latent object parts. Hence, metric learning ensures that the deep representation of images showing similar object classes which share some parts are closer than the representations of images which do not have common parts. Our results show that we significantly outperform the state of the art, as well as get similar performance to the common episodic training for fully-supervised few-shot learning on the Mini-Imagenet and Tiered-Imagenet datasets.
- Keywords: Unsupervised Few-shot Learning, Deep Metric Learning, Self-supervised Learning