Contents

Correspondence Networks with Adaptive Neighbourhood Consensus

该文提出了一种建立两张图像间密集语义对应关系的模型(ANC-Net),

  • non-isotrppic(非各向同性) 4D convolution kernel – 核心,
  • multi-scale self-similarity module
  • orthogonal loss

ANC-Net 以两张图像作为输入,输出为4D correlation map – 包含两张图像间所有可能匹配的匹配分数。

https://gitee.com/shilongshen/image-bad/raw/master/img/20200910101415.png

https://gitee.com/shilongshen/image-bad/raw/master/img/20200910101440.png

Method

https://gitee.com/shilongshen/image-bad/raw/master/img/20200910105234.png

网络结构如图所示。

  • 输入(Is,It)(I^s,I^t)
  • feature extractor F\mathcal{F} ->输出 FsF^sFtF^t
  • multi-scale self-similarity S\mathcal{S} ->输出 multi-scale self-similarity SsS^sStS^t->captures the complex self-similarity feature map
  • We can then obtain the 4D correlation map CsC_s from SsS^s and StS^t , and the 4D correlation map CfC_f from FsF^s and FtF^t . However, CsC_s and CfC_f are often noisy as they lack the constraints to enforce the correspondence validity, and thus are unreliable for directly extracting correspondences.
  • ANC module N\mathcal{N} -> a stack of non-isotropic 4D convolutions ->takes CsC_s and CfC_f as inputs , refining them by considering neighbourhoods with varying sizes.
  • Finally, the ANC module combines the refined correlation maps by simply summing up the two, producing a single 4D correlation map Cˉ\bar{C} from which reliable correspondences can be retrieved.

https://gitee.com/shilongshen/image-bad/raw/master/img/20200910111121.png

https://gitee.com/shilongshen/image-bad/raw/master/img/20200910112010.png

给定特征FRhf×ωf×dF\in \mathbb{R}^{h_f\times \omega_f \times d} , self-similarity map 度量每一个特征位置之间的局部相似性。

计算FF中位置(i,j)(i,j)处的特征fi,jf_{i,j}的self - similarity map 的方法是计算与它自身和它邻域之间的余弦相似性。

如上图所示,考虑一个特征fi,jf_{i,j}3×33\times 3邻域,可以计算得到3×33\times 3个self-similarity scores,将其进行向量化,最终可以得到self-similarity features map S0Rhf×ωf×9S_0\in \mathbb{R}^{h_f \times \omega_f \times 9}

为了进一步得到不同self-similarity features 之间的相关性,对S0S_0进行两次zero-padding convolution 操作,分别得到S1S_1S2S_2。然后将三个不同尺寸的feature map S0,S1,S2S_0, S_1 ,S_2进行叠加得到SS,其作为最后一层的输入。

https://gitee.com/shilongshen/image-bad/raw/master/img/20200910171240.png

什么是4D卷积核?

各向同性的4D卷积核被用于修正4D correlation map 。各向同性的4D卷积核可以被认为建立了两张图像相同尺寸的两个邻域。

然而在实际的图像中的物体,有变化的尺寸和形状,描述相同语义的两个邻域可能有不同的尺寸。因此使用相同尺寸的邻域可能会引入噪声(例如不相关的背景)。为了解决这一问题引入了ANC module ,其包含了一系列的非各向同性4D卷积核。

在模型中,为了解决物体变化的尺寸和形状,将各向同性4D卷积核与非各向同性4D卷积核相结合,使得模型能够动态的决定使用哪一个卷积核

https://gitee.com/shilongshen/image-bad/raw/master/img/20200910185933.png

https://gitee.com/shilongshen/image-bad/raw/master/img/20200910190152.png

https://gitee.com/shilongshen/image-bad/raw/master/img/20200910190233.png

https://gitee.com/shilongshen/image-bad/raw/master/img/20200910191827.png

https://gitee.com/shilongshen/image-bad/raw/master/img/20200910191922.png