Supporting ground-truth annotation of image datasets using clustering

As more subject-specific image datasets (medical images, birds, etc) become available, high quality labels associated with these datasets are essential for building statistical models and method evaluation. Obtaining these annotations is a time-consuming and thus a costly business. We propose a clustering method to support this annotation task, making the task easier and more efficient to perform for users. In this paper, we provide a framework to illustrate how a clustering method can support the annotation task. A large reduction in both the time to annotate images and number of mouse clicks needed for the annotation is achieved. By investigating the quality of the annotation, we show that this framework is affected by the particular clustering method used. This, however, does not have a large influence on the overall accuracy and disappears if the data is annotated by multiple persons.