Multi-Scale Correlation for Deep Homography Estimation

Abstract

In this paper, we propose a novel multi-scale correlation network (MSCNet) for homography estimation from coarse to fine. First, we extract multi-scale features through a siamese network to generate global and local correlations from feature maps of different scales. Second, we use a group dilated deconvolution block to capture global mapping by increasing the receptive fields in terms of different levels. Third, we employ the channel and spatial attention mechanism to achieve local refinement for small displacements. Finally, we adopt a knowledge distillation strategy to lightweight our model while maintaining relatively high estimation performance. Experimental results on Microsoft Common Objects in Context (MSCOCO) dataset show that our proposed MSCNet outperforms the state-of-the-art approaches in terms of accuracy and parameter count.