.png){kind=icon}
.png){kind=icon}
Return
Learning-based correspondence classifier with self-attention hierarchical network
.png){kind=icon}
PRE
AI
DOI：10.1007/s10489-023-04789-w.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Finding valid correspondences is of considerable significance to image matching, which has been regarded as the key of numerous vision-based tasks. Current methods usually have drawbacks in sets with high proportion of outliers. To address the problem, given a set of putative correspondences in two images, this paper proposes a novel framework (named SAH-Net) to remove outliers and recover camera pose through essential matrix using an end-to-end network. The proposed SAH-Net is hierarchical with a multi-scale structure, which consists of correspondence level and cluster level. First, correspondence level takes advantage of two-view geometry to learn correspondence features. Next, in order to integrate structural information of the scene, correspondences are pooled via a self-attention method. Additionally, SAH-Net applies a spatial correlation operation after the clustering, separating features into segments and learning the spatial characteristics of clustered nodes. Finally, clusters have been integrated with spatial information, and they are recovered to original scale via a learned upsampling operation. Extensive experiments are conducted on remote sensing image registration, general image matching (outdoor and indoor image datasets respectively) and loop closure detection, which demonstrate the excellence of SAH-Net in mismatch removal and relative pose estimation compared to other state-of-the-art competitors.
Keywords:
Classifier
Mismatch removal
Outlier
Pose estimation
Self-attention