A dynamic label assignment strategy for one-stage detectors

In object detection field, label assignment (LA) is an important step in determining the detection accuracy, which assigns positive and negative labels for the training samples, so that the prediction loss could be calculated. Therefore, how to realize a more reasonable LA has always been a major concern for computer vision experts. Considering the difficulty of current LA strategy in adapting to different scenarios and the lack of interaction between the classification and localization tasks. We propose a novel dynamic LA scheme for one -stage object detector. Firstly, the qualities of the anchor boxes are computed based on the outputs of both classification and localization, which are used to assign the positive and negative samples and will also be adjusted during training. Secondly, the positive and negative samples for the classification and localization tasks are decoupled, and independent LA strategies are developed for each task. Finally, the interaction between the two network heads are enhanced through multiple shared convolution blocks so as to complete the two tasks in a more collaborative manner. Extensive experiments conducted on MS COCO, PASCAL VOC and CrowdHuman to support our design and analysis. With the newly introduced LA strategy, we improve the detection accuracy of existing one -stage detector to a new level.