Multi-Grid Homography Estimation Based on Swin SC-Mixer

Homography estimation is an important and fundamental work in computer vision, and the accuracy of homography estimation has a very important impact on downstream tasks. Traditional homography methods are limited by weak texture images and are not robust enough in practical applications. As an alternative method, CNN can have better robustness. However, CNN cannot obtain global attention due to its limited receptive field, so the accuracy will be limited in the homography estimation task. This paper proposes a new model called Swin SC-Mixer. Swin SC-Mixer uses a shifted window mechanism to make the size of the model smaller, and the computational complexity is linear with the image. Applying the Swin SC-Mixer model to multi-grid homography estimation achieves higher accuracy than CNN-based and Transformer-based methods. And unlike previous methods, this paper estimates a multi-grid homography instead of single homography, so images with a certain parallax can be aligned. Experiments show that our method is fast and effective.