cascade_s2anet.md

Cascade S²A-Net

Cascade S²A-Net (CS²A-Net) can be regarded as a generalized S²A-Net. It consists of several detection heads, and each head contains an Alignment Convolution and two subnetworks, .i.e, cls and reg. The overall archetecture of CS²A-Net is shown below.

The difference between CS²A-Net and S²A-Net is:

• CS²A-Net head aligns features first by an Alignment Convolution Layer (ACL), and produce cls_score and bbox_pred later. While the FAM in S²A-Net does clssification and regression first, then produces aligned features.

• CS²A-Net aligns features in each stage, while feature alignment in S²A-Net is only appeared at the end of FAM. So even the one stage version of CS²A-Net, .i.e, CS²A-Net-1s can produce aligned features.