DSLA: Dynamic smooth label assignment for efficient anchor-free object detection

Anchor-free detectors basically formulate object detection as dense classification and regression. For popular anchor-free detectors, it is common to introduce an individual prediction branch estimate the quality of localization. The following inconsistencies are observed when we delve into practices estimation. Firstly, for some adjacent samples which assigned completely different labels, trained model would produce similar scores. This violates training objective leads performance degradation. Secondly, found that detected bounding boxes with higher confidences contrarily have smaller overlaps corresponding ground-truth. Accurately localized be suppressed by less accurate ones in Non-Maximum Suppression (NMS) procedure. To address inconsistency problems, Dynamic Smooth Label Assignment (DSLA) method proposed. Based on concept centerness originally developed FCOS, a smooth assignment strategy label smoothed continuous value [0,1] make steady transition between positive negative samples. Intersection-of-Union (IoU) predicted dynamically during coupled label. dynamic supervise branch. Under such supervision, estimation naturally merged branch, simplifies architecture detector. Comprehensive experiments conducted MS COCO benchmark. It demonstrated that, DSLA can significantly boost accuracy alleviating above detectors. Our codes released at https://github.com/YonghaoHe/DSLA.