Correspondence Learning via Correspondence Embedded and Channel Recalibration Network

Abstract

Correspondence learning is pivotal to many computer vision-based tasks. Existing methods regard each correspondence equally along the channel dimension, which weakens the feature representation capability of the network. To alleviate this problem, we propose a Correspondence Embedded and Channel Recalibration Network, named CECR-Net, to predict the inlier probability of each correspondence and recover camera poses. The proposed CECR-Net is designed to explore the potential impact of correspondences on the channel dimension, and recalibrate the weight of each channel, so that our CECRNet can capture more exact contextual information. Experiments show that our CECR-Net is effective in outlier removal and camera pose estimation tasks on challenging public datasets.