Abstract
Aerial imagery objects present characteristics such as arbitrary orientations, dense distribution, and a wide range of scales. Mainstream detectors add angles to traditional rectangular bounding boxes to better fit the object boundaries, but this makes the detection task even more challenging. Non-Maximum Suppression (NMS) is a filter designed to address duplicate detections in dense detectors, extracting the most representative output from multiple detection results for a single object. However, the NMS algorithm has fundamental problems, including the risk of wrong suppression due to confidence-based prioritization, poor computational speed from sequential processing, and inefficient use of its output data. To address the above challenges, and based on the data patterns of objects in remote sensing imagery, we reviewed NMS as a local density peak clustering task and proposed a novel post-processing method called DPNMS. We integrated information across each output through statistical analysis to construct its density representations, and then searched for peak values within valid neighborhoods defined by the IoU threshold. To prevent suppressing effective boxes, we designed a “select–release” cyclic workflow to correct the chosen result. The proposed method integrates the detector’s output information to achieve competitive results. Our method does not require sorting the localization information in reverse order of confidence, and each operation is independent of the others, allowing for parallelization and resulting in high computational efficiency. Extensive experiments conducted on various datasets and multiple platforms have validated the effectiveness and stability of the proposed method. The unique filtering strategy offers a new framework for post-processing methods in remote sensing image object detection.