Prime Label Learning from Multi-label Aerial Image: A Novel Weakly Supervised Task

In the task of multi-label aerial image classification, various objects and land cover in an image are usually represented by multiple labels which are treated equally. However, from a semantic point of view, the importance of multiple labels are different in a specific scene. There is often a prime label in the image that plays a "leading" role. Obtaining the most important label from candidate multiple labels is crucial, because it best represents the semantics of the entire image. In this letter, we attempt to automatically obtain the prime label of each image from several existing multi-label aerial image datasets without additional supervision cost. In other words, the prime labels are only used to evaluate the performance of models during testing and do not participate in the training process. Therefore, it is essentially a weakly supervised learning task. For this novel aerial image classification task, corresponding datasets are provided in this letter firstly, including over head images with multi-labels for training and prime labels for testing. Then the baselines on the above datasets are provided. Finally, a new prime label learning method is proposed, which improves the baseline accuracy by about 14% and reaches the state-of-the-art on current datasets.