Each pair of rain and object patches are visually similar. Wecompute thedifference (L1 distance) between the average kernels of each pair of confusing rain and object patches. We accumulate and average the differences of all pairs. We change the number of sampled kernels for computing the average kernel. In Figure 13(b), we report the difference of average kernels of the confusing patch pairs.

The feature masking contains two convolutional layers. It computes the rain (or object) feature map. We use Adam solver to optimize the parameters of SEIDNet. The performances are reported on the test set of Rain100H. We report the results in Table 1 (also see Table 1 of the main paper).