Boosting Dense Long-Tailed Object Detection from Data-Centric View

Abstract

AbstractSeveral re-sampling and re-weighting approaches have been proposed in recent literature to address long-tailed object detection. However, state-of-the-art approaches still struggle on the rare class. From data-centric view, this is due to few training data of the rare class and data imbalance. Some data augmentations which could generate more training data perform well in general object detection, while they are hardly leveraged in long-tailed object detection. We reveal that the real culprit lies in the fact that data imbalance has not been alleviated or even intensified. In this paper, we propose REDet: a rare data centric detection framework which could simultaneously generate training data of the rare class and deal with data imbalance. Our REDet contains data operations at two levels. At the instance-level, Copy-Move data augmentation could independently rebalance the number of instances of different classes according to their rarity. Specifically, we copy instances of the rare class in an image and then move them to other locations in the same image. At the anchor-level, to generate more supervision for the rare class within a reasonable range, we propose Long-Tailed Training Sample Selection (LTTSS) to dynamically determine the corresponding positive samples for each instance based on the rarity of the class. Comprehensive experiments performed on the challenging LVIS v1 dataset demonstrate the effectiveness of our proposed approach. We achieve an overall 30.2% AP and obtain significant performance improvements on the rare class.