Explicit homography estimation improves contrastive self-supervised learning