*4.1. Quantitative Results*

Tables 2 and 3 are the quantitative results on the SSDD and HRSID datasets. From Tables 2 and 3, GCBANet achieves the best accuracy, that is, on SSDD, the box AP reaches 68.4% and the mask AP reaches 63.1%; on HRSIS, the box AP reaches 69.4% and the mask AP reaches 57.3%. The mask prediction has poorer indexes than the box prediction, because the pixel-level segmentation task is more difficult than the box-level detection task. GCBANet outperforms the other nine competitive models by a significant degree. On SSDD, it achieves 2.8% higher box AP and 3.5% higher mask AP than the previous most advanced model; on HRSID, they are 2.7% and 1.9%. This fully reveals the state-of-the-art performance of GCBANet. This accuracy advantage benefits from the combined action of the proposed GCIM-Block and BABP-Block. Certainly, the speed of GCBANet does not win advantages, compared with others, thereby the speed optimization is required in the future. Moreover, although YOLACT [68] offers the fastest detection speed because it is a one-stage model, its accuracy is too poor to satisfy application requirements.

Table 1 shows the computational complexity calculations of different methods. Here, we adopt the floating point of operations (FLOPs) to measure calculations whose unit is the giga multiply add calculations (GMACs) [69]. From Table 1, the calculation amount of GCBANet is more than the others, so future model computational complexity optimization is needed.

**Table 1.** Computational complexity calculations of different methods. Here, we adopt the floating point of operations (FLOPs) to measure calculations whose unit is the giga multiply add calculations (GMACs) [69].




**Table 3.** Quantitative results on HRSID. The suboptimal method is marked by Underline "—".
