全局光度对齐(Global Photometric Alignment):由于全局域偏移主要与低级图像属性有关,我们的工作中提出了全局光度对齐方法,将目标域的低级图像属性转移到源域图像中。观察到在不同场景中,图像的空间亮度分布可能非常复杂。还需要注意的是,直接操作 RGB 通道可能会导致严重的伪影和假颜色。相比之下,a 和 b 色彩通道的空间色彩分布总是具有类似的钟形直方图。
因此,我们对亮度和颜色采用不同的处理方式:我们通过仅在颜色通道 a 和 b 上执行经典的直方图匹配,而在亮度通道 L 上运用 gamma 校准的方式将源域图像与目标域参考图像进行匹配,以避免引入常见的直方图匹配结果中的伪影。下图展示了全局光度对齐的具体流程:(a)输入源域图像和(b)随机选择的目标域图像在(c)Lab通道上对齐,生成(d)对齐后的图像。
全局流形对齐(Global Manifold Alignment):诸如局部线性嵌入(LLE)和Isomap等方法通常用于描述流形,但它们对基于梯度反向传播的训练来说计算成本太高。在这里,我们使用 K - 均值算法简化计算。由于 LLE 使用分段线性模型来逼近高维特征流形,K - 均值可以被视为流形的分段常数逼近。通过 K - 均值获得的每个质心都是局部区域的常数逼近。
[1] Zhu J Y, Park T, Isola P, et al. Unpaired image-to-image translation using cycle-consistent adversarial networks[C]//Proceedings of the IEEE international conference on computer vision. 2017: 2223-2232.[2] Yang Y, Soatto S. Fda: Fourier domain adaptation for semantic segmentation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 4085-4095.[3] Wu Z, Han X, Lin Y L, et al. Dcan: Dual channel-wise alignment networks for unsupervised scene adaptation[C]//Proceedings of the European Conference on Computer Vision (ECCV). 2018: 518-534.[4] Pan F, Shin I, Rameau F, et al. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 3764-3773.[5] Li Y, Yuan L, Vasconcelos N. Bidirectional learning for domain adaptation of semantic segmentation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019: 6936-6945.[6] Wang H, Shen T, Zhang W, et al. Classes matter: A fine-grained adversarial approach to cross-domain semantic segmentation[C]//Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIV. Cham: Springer International Publishing, 2020: 642-659.[7] Zhang Q, Zhang J, Liu W, et al. Category anchor-guided unsupervised domain adaptation for semantic segmentation[J]. Advances in neural information processing systems, 2019, 32.[8] Wang Z, Yu M, Wei Y, et al. Differential treatment for stuff and things: A simple unsupervised domain adaptation method for semantic segmentation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 12635-12644.