Question about warping keypoints

[hkkkpang]

From here we know that the source image is the result of homography H applied to target image. We can also know that the source information in model part is from augmented image and target information is from original image by this part. This line would transform the source keypoints to target keypoints by using homography H. However, since the H is used to transform the target image to source image based on the code I mentioned in the first sentence. Should we use the inverse of homography to transform the source keypoints to target keypoints instead of homography itself? I am really confused about this problem.