Since the advent of deep learning, deep learningbased homography estimation methods have been continuously

proposed. However, these methods often require a fixed input size (e.g., 128x128) and are computationally inefficient to apply to larger images. These constraints hinder their deployment in real-world applications, particularly on mobile devices with limited computing power. In this paper, we propose an efficient method to estimate the global homography of a large image. Our approach leverages a pre-trained homography network on small 128x128 patches. During inference, we extract four patches from the image quadrants, estimate their respective 4-point offsets, and then combine the single most relevant corner offset from each patch to form the final global homography. Experimental results demonstrate that our method achieves state-of-the-art performance, maintaining a consistently low error regardless of the input image size, unlike baseline methods whose errors increase with resolution.