基于深度阴影特征增强的任意至任意重光照

任意至任意重光照利用隐含在引导图像中的光照重新照明源图像。现有的任意至任意重光照方法由于采用端到端的学习方式，导致阴影特征与色温特征高度耦合，进一步影响了阴影生成的准确性。为此，本文提出了一个基于深度阴影特征增强的任意至任意重光照方法。该方法的关键是设计一个额外的阴影解码器，从隐式表征中直接生成对应的阴影图像。同时，为了充分利用学习到的阴影特征，我们引入一个基于注意力机制的特征融合模块，实现重光照特征与阴影特征的自适应融合。另外，我们实验性地发现，利用多项式核函数把源图像映射到高维特征后，再作为网络输入，能进一步提升性能。在VIDIT数据集上的实验表明了本文所提方法的有效性。 

Enhancing Deep Shadow Features for Any-to-Any Relighting

? ?Any-to-any relighting is to relight the source image with the illumination implicitly given in the guide image. Existing any-to-any relighting methods adopt an end-to-end learning way， resulting in a high coupling between shadow features and color temperature features， which further affects the accuracy of shadow generation. To this end， this paper proposes an any-to-any relighting method based on deep shadow features enhancement. The key to this method is to design an additional shadow decoder to directly generate the corresponding shadow image from the implicit representations. At the same time， to make full use of the learned shadow features， we introduce a feature fusion module based on the attention mechanism to realize the adaptive fusion of relighting features and shadow features. In addition， we experimentally found that using a polynomial kernel function to map the source image to high-dimensional features and then taking them as network input can further improve the performance. Experiments on the VIDIT dataset demonstrate the effectiveness of the proposed method.