5.1.2. Component Analysis in GCIM-Block

We also make a component analysis in GCIM-Block, as shown in Table 5. From Table 5, each component can offer an observable accuracy improvement, either the box AP or the mask AP. This indicates that our well-designed idea is reasonable and our theoretical analysis in Section 2.1 is correct. Moreover, we observe that GFSA does not improve the box prediction performance, but it improves the mask prediction performance further. This is because the global feature self-attention is pixel-sensitive, which can enable better pixel classification capability.

**Table 5.** Quantitative Results Component Analysis in GCIM-Block.


1 CAFR denotes the content-aware feature reassembly. 2 MRFFR denotes the multi receptive-field feature response. 3 GFSA denotes the global feature self-attention.
