Search

• We propose a new learning-based oversegmentation method for 3D meshes. • We develop a segmentation-aware loss that guides the network in generating boundary-aware superfacets. • We exploit the hierarchical clustering of our method for soft face-superfacet association. • We exploit quantitative and qualitative evaluations of the quality of our oversegmentation method. Superfacets are generated by clustering adjacent mesh faces that share similar characteristics, which can serve as processing units in downstream mesh applications. While there are existing deep neural networks that generate superpixels and superpoints/supervoxels from images and point clouds respectively, the current oversegmentation methods in 3D meshes mostly rely on hand-crafted features that are extracted using non-differentiable algorithms to generate superfacets. Nevertheless, these methods cannot leverage the feature extraction abilities of deep neural networks to generate superfacets in an end-to-end fashion. Therefore, we propose an end-to-end trainable deep neural network that learns to generate boundary-aware superfacets from 3D meshes. Specifically, our network learns a soft face-superfacet association map from faces and their adjacency relationships. Moreover, we develop a segmentation-aware loss on the faces that train the network to predict their labels in an end-to-end manner. We evaluate the performance of our method using mesh adaptations of two well-known superpixel evaluation metrics where experimental results demonstrate that the performance of our proposed network surpasses that of other state-of-the-art mesh oversegmentation methods, and in doing so simultaneously improves superfacet-based semantic segmentation. [ABSTRACT FROM AUTHOR]

• Point cloud and Mesh models typically use cosine similarity without normalization. • Softmax loss devise a lower-bound and cause a convergence problem. • The softmax loss function is reformulated to let the network converge properly. • Reformulation of softmax loss shows performance improvements over previous methods. Feature normalization has been a crucial step in convolutional neural networks (CNNs) in the past few years. Discriminative feature abstraction is indispensable for boosting the overall performance of learning models. For 3D data, in both point cloud and mesh models, the inner product (cosine similarity) is frequently applied for similarity estimation without normalization. In this paper, we first demonstrate that softmax loss will produce a lower bound and cause a convergence problem if we directly perform feature comparison by cosine similarity. Second, we propose a revised formulation of the softmax loss function that uses an optimized scaled cosine similarity with normalization in the training phase rather than the inner product of two vectors, allowing the network to converge correctly. Experiments show significant improvements over previous methods on several datasets. The SHREC dataset with split16 and split10 achieved 99.16% and 97.66% accuracy, respectively, and the Cube dataset achieved 96.96% accuracy, surpassing previous methods in a mesh classification task. [ABSTRACT FROM AUTHOR]