Motivation:Notice the variability across different object instances. Any weakly or un- supervised technique for part segmentation needs to reason about correspondences between different images which is challenging in such diverse image collections.
Method:
一个好的part segmentation要具备如下特点:
Geometric concentration: Parts are concentrated geometrically and form connected components.
Robustness to variations: Part segments are robust with respect to object deformations due to pose changes as well as camera and viewpoint changes.
Semantic consistency: Part segments should be seman- tically consistent across different object instances with appearance and pose variations.
Objects as union of parts: Parts appear on objects (not background) and the union of parts forms an object.
Overall:
没有使用任何监督信息,得到热图后通过loss进行一系列约束的自监督算法,实现细节如下:
--> 通过CNN+softmax 得到的part response map R|(K+1)*H*W
--> 归一化 R^ 并将背景设置为常数0.1 目的:增强weak part responses
-->argmax得到结果
-->不引入任何监督信息,仅可微分loss对结果进行约束
-->test:用单张图片
1) Geometric Concentration Loss
缘由:the same object part are usually spatially concentrated within an image and form a con- nected component unless there are occlusions or multiple instances. 同类别集中出现,除非有遮挡。
鼓励一张图上同类别的所有点,尽可能接近part中心。
励零件的几何集中,并试图最小化空间概率分布函数 R(k, u, v)/zk 的方差。
2) Equivariance Loss 等变性
part segmentation should be robust to the appearance and pose variations。
3) Semantic Consistency Loss
等变性2)损失有利于对某些对象变化具有鲁棒性的部分分割,但这些转换不足以在不同实例之间产生一致性,因为图像之间的外观和姿势变化太高而无法通过任何人工转换建模 。 为了鼓励不同对象实例之间的语义一致性,我们需要在我们的损失函数中明确利用不同的实例。