5.2.1. Effectiveness of BABP-Block

Table 6 shows the quantitative results with and without BABP-Block. BABP-Block offers a 2.8% box AP gain and a 1.8% mask AP gain, showing its effectiveness. It can offer better box prediction by predicting the boundary so as to enable better mask prediction. Thus, this boundary estimation scheme should be more suitable for SAR ships.

**Table 6.** Quantitative Results with and Without BABP-Block.


5.2.2. Component Analysis in BABP-Block

We also make a component analysis in BABP-Block as shown in Table 7. From Table 7, each component is conducive to boosting accuracy, which shows their effectiveness. BAFE is able to extract more boundary-sensitive features so as to ensure accurate boundary bucketing coarse localization. BBCL locates four sides of the box to avoid long-distance regression, which boosts information flow. BRFL can enable more refined box regression. Finally, BGCR can leverage the boundary reliability to guide classification scores, so as to screen the detection results again, leading to more reliable predictions. As a result, the accuracy is improved progressively.

**Table 7.** Quantitative Results Component Analysis in BABP-Block.


1 BAFF denotes the boundary-aware feature extraction. 2 BBCL denotes the boundary bucketing coarse localization. 3 BRFL denotes the boundary regression fine localization. 4 BGCR denotes the boundary-guided classification rescoring.
