作者指出,在"there exists displacements with large variance between the anchors and the targets"(即anchor与目标之间存在大量位置偏移时),传统的中心和size预测并不能很好地解决问题。因此引入side-aware预测。
然而在motivation中,对于方法的提出,作者却给出了几乎没有说服力的解释:
Empirically, we observe that when we manually annotate a bounding box for an object, it is often much easier to align each side of the box to the object boundary than to move the box as a whole while tuning the size.
method:
对于二阶段检测器的整体流程图:
Side-Aware Feature Extraction
在RoIAlign之后,经过2个3 × \times × 3卷积得到F。再用自注意力机制提高RoI特征 : 使用1 × \times × 1卷积操作后分别在Y,X轴做softmax得到两种不同的attention map M