MAIM: a mixer MLP architecture for image matching

Zhiwei Shen, Bin Kong, Xiaoyu Dong

Abstract: Recent advances in multilayer perceptron (MLP) models have provided new effective network architecture designs for computer vision tasks. Compared with convolutional neural networks (CNNs) and visual transformers, MLP-based visual backbones have less induction bias, which can improve the sample utilization efficiency and reduce computational costs. Therefore, we designed the Mixer MLP Architecture for Image-Matching (MAIM), which is a coarse to fine-level detector-free image-matching scheme. Accordingly, we constructed a mixer MLP architecture called Mixer-WMLP, which evenly divides the feature map into non-overlapping windows, spreads each window as a token, achieves the exchange of token information between spatial locations, channels features through a two-layer MLP structure in the coarse-level model, and then feeds the windows with dense fine-level matching, thereby producing the final matches. Furthermore, the implemented global field-of-view mixer MLP framework for image-matching incurs a low computational cost. By conducting experiments with indoor and outdoor relative poses, our MLP architecture is compared with CNN and transformer-based image-matching methods. Our method has significant advantages in terms of real-time performance and largely reduces computational cost, proving its effectiveness in image-matching tasks.