Few-Shot Open World Learner

Abstract Computer vision based recognition systems in dynamically changing environments require continuously updating datasets with novel detected categories while maintaining equally high performance on previously established classes. These requirements can be addressed by the concept of Open World Learning introduced by (Chen et al., 2018). We propose a novel framework as a solution to the open world learning problem. This solution is based on the few-shot classification strategy in combination with an outlier detection module. The few-shot classifiers utilized in this work provide a similarity-based classification scheme and present an excellent solution to the incremental learning problem natively, requiring no training and only a few labeled examples of new classes before adapting to them. The discovery of new classes is performed by an outlier detection module that utilizes the similarity space created by the few-shot classifier to identify sufficiently different samples from the known classes and remove them from the classification process. This paper highlights the best combination of few-shot and outlier detection algorithms. Extensive experiments with different combinations of these algorithms are conducted in the pursuit of the ideal components of the framework. Results of the experiments show that we are successful in creating a novel implementation of an open world learner with a very limited loss of accuracy compared to the base few-shot algorithm.