homography is the wrong approach.
the depth view is taken from a different position, not the same as the visual view.
you need to project the visual view onto the point cloud/mesh, taking the transform between views into account.
their website looks like they have examples for that.