EdgeNet: Deep metric learning for 3D shapes

We introduce EdgeNet, a metric learning architecture for extracting semantic local shape features, directly applicable to a wide range of shape analysis applications such as point matching, object classification, shape segmentation, and partial registration. EdgeNet is based on a novel technique to keep edge-wise correspondences in the deep feature space and encodes the local structure into the learned features. It is trained under the supervision of edge-wise correspondences by using the 3D coordinates. The training loss combines a bi-triplet loss to enforce feature variations between the semantic matching points in the feature space, a transformation loss to encourage consistency between corresponding edges after alignment transformation, and a smoothness loss guarantees the flatness between the nearest points in the feature space. The learned features are proved to encode local content, structure, and asymmetry for 3D shapes. Our network can be adapted to either 3D meshes or point clouds. We compare the performance of the EdgeNet with existing state-of-the-art approaches and demonstrate the efficiency and efficacy of EdgeNet in three shape analysis tasks, including shape segmentation, partial matching, and shape retrieval.