Adaptive spatial and scale label assignment for anchor-free object detection

Abstract

In recent years, anchor-free object detection has attracted widespread attention due to its simplicity and efficiency. The mainstream anchor-free object detectors allocate positive/negative candidate samples through prior guidance at a fixed spatial position and assign positive/negative samples according to predefined scale constraints. However, artificially designing assignment strategies according to prior data distribution may hinder further optimization of label assignment. To this end, this paper proposes Adaptive Spatial and Scale Label Assignment (ASS-LA) to improve the performance of anchor-free object detection. Positive/negative samples are distributed from different pyramid levels using spatial and scale constraints. Specifically, an adaptive Intersection-over-Union (IoU) space assignment is designed to select candidate positive sample points. The membership degree is introduced at each pyramid level to adaptively fuzzy the scale assignment range so that the detector selects the final positive sample from the candidate sample points. Furthermore, a reference box is introduced to design the predicted IoU branch of coupled regression. In the inference stage, the predicted IoU and classification scores are combined as the confidence of the regression bounding box to alleviate the inconsistency between classification and regression. Extensive experiments show that our method achieves comparable performance to other existing label assignment schemes. With the introduction of ASS-LA, the anchor-free object detector has significant performance improvements without introducing other overhead.