<p>Label assignment strategy plays a vital role in supervised object detection systems. However, the handcrafted label assignment used in most previous studies is usually suboptimal because the assignment strategy is not adaptive to the training objective. Considering the limitations of handcrafted label assignments, we propose a novel object detection framework consisting of a triple-head network and a joint-loss-aware label assignment strategy. The foregrounds and backgrounds are first determined based on a ground truth (GT) box. A label assignment strategy was proposed, which leveraged the joint loss of classification and regression as a unified sampling criterion to distinguish between positive and negative samples, while dynamically assigning optimal GT label for supervised learning. In addition, ambiguous samples from overlapping areas of multiple GT boxes were sampled based on a loss-weighting mechanism. We added a pre-regression branch to the head of FCOS, forming a triple-head detection network alongside the original classification and regression branches. The pre-regression branch performed coarse localization prediction on the object, provided additional contextual information for the classification and regression branches. Extensive experiments showed that the proposed method can effectively improve the detection performance for small and medium objects. And the proposed method achieved an average precision (AP) of 45.0% on the MS COCO test-Dev2017 dataset, achieving performance comparable to state-of-the-art methods.</p>