.png){kind=icon}
.png){kind=icon}
Return
Generalized unsupervised functional map learning for dense correspondence
.png){kind=icon}
PRE
AI
DOI：10.1007/s00371-022-02752-3.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Inspired by deep functional map methods, we present a generalized unsupervised functional map learning approach for arbitrary 3D shape correspondence. Unlike prior methods, they either require extensive data training or rely on input features; our model directly operates on point clouds and learns both deep features and optimized basis function without the constraint of geometric connectivity and the assumption of isometry. We propose a novel scheme that combines structural embedding based on Mahalanobis distance and locally linear embedding to learn the optimized feature basis. Furthermore, the constructed shape descriptors effectively optimize the estimation of functional map and dense correspondence through a tri-level regularization mechanism that enforces penalties on global structural properties, representation error and pair-wise Mahalanobis distance distortion, which significantly improves the performance of unsupervised learning. Extensive experiments in shape matching show that our method can learn from less training data and has better generalization ability compared with the state-of-the-art supervised and unsupervised methods.
Keywords:
Deep learning
Shape correspondence
Functional map
Unsupervised learning
Point clouds
Journal
IF:
2.9
Papers:
4.6K
Citations:
6.5K
