Dual-Threshold: A Hierarchical Label Assignment Scheme Considering Scale Information For Small Object Detection

Small object detection has always been a focal point and challenge within the object detection task. Despite the availability of various advanced object detectors, the performance of detecting small objects in general scenes remains unsatisfactory. A crucial factor contributing to this is the imbalance of samples between small objects and objects with other scales within general scene datasets. In order to address this issue, we approach it from the perspective of label assignment. Specifically, we distinguish between small objects and objects with other scales, employing a dual-threshold approach based on pseudo anchors. By setting different threshold criteria for small objects and objects with other scales after generating pseudo anchors during label assignment process, we increase the sample ratio of small objects without any additional computational cost or extra parameters for training and testing. By incorporating dual-threshold metric into an anchor-free object detection framework, we observe consistent improvements of the proposed method for small objects on general object detection benchmarks such as PASCAL VOC and MSCOCO. Our method achieves 40.6% AP on MSCOCO and reaches 24.0% AP_s, which is 0.4% higher than the anchor-free baseline and behaves better than other recent state-of-the-art detectors.