Small Object Localization with 90% Annotation Reduction by Positive-Unlabeled Learning

Small object localization is one of the most challenging tasks owing to the poor visual appearance and noisy representation caused by the intrinsic structure of small targets. Recent advances in localizing small objects are mainly dependent on regression-based counting approaches, which require considerable annotations for training. As a contrast, human learners can quickly master labeling skills from only a few annotation examples. In this paper, we attempt to simulate this training mechanism and propose a novel positive-unlabeled (PU) learning based approach that can localize small objects by learning from partial point annotations. We evaluate our approach on five typical datasets of small objects involving a single cell, an animal/insect, and human crowds. Quantitative experimental results show that our approach has achieved inspiring localization performance (F1 score > 0.75) even under the supervision of less than 10% of the overall point annotations. This approach paves the way for low-annotation-cost single-cell analysis within microfluidic droplets.