Open-set low-shot classification leveraging the power of instance discrimination

Abstract

In search, exploration, and reconnaissance operations of autonomous ground vehicles, an image recognition capability is needed for specifically classifying targeted objects (relevant classes) and at the same time identifying as unknown (irrelevant and unseen) objects that do not belong to any known classes, as opposed to falsely classifying them in one of the relevant classes. This paper integrates an unsupervised learning feature extraction framework based on the Instance Discrimination method with an Open-Set Low-Shot (IDLS) classifier for creating the desired new capability. Unlabeled images from the vehicle’s operating environment are used for training the feature extractor while a modest number (less than 40) images for each relevant class and unlabeled irrelevant images are used for training the Open-Set Low-Shot (OSLS) classifier in a manner that enables recognition of images unseen during training as irrelevant. The value and the accuracy of the new IDLS approach are demonstrated through a thorough comparison with alternative unsupervised and fully supervised methods.