MeshCL: Towards robust 3D mesh analysis via contrastive learning

Abstract

Recent years have witnessed the success of deep learning in 3D mesh analysis, including classification and segmentation. Previous works mainly rely on supervised learning, which trains models on the labeled 3D mesh data with the cross-entropy loss. However, in the supervised training, the very limited amount of 3D mesh data restricts the feature’s discriminative power and the model’s robustness. To address or at least alleviate this issue, we propose to regularize the learned representations by associating the data points in feature space via contrastive learning. To achieve this, we first explore mesh samples with diverse geometric structures, and then apply contrastive regularization into the positive and negative pairs sampled from the feature space. In this way, our method, MeshCL, can learn robust feature representations invariant to the geometric deformation for 3D manifold mesh. In detail, to augment the geometric diversity, we examine different data augmentation techniques for mesh data. Furthermore, considering the geometric characteristics of 3D mesh, we develop more efficient positive/negative pair sampling strategies for the dense prediction task. To verify the effectiveness of MeshCL, we study two typical mesh analysis tasks, classification and segmentation. Experimental results demonstrate that MeshCL not only achieves state-of-the-art performance on the two tasks mentioned above but also significantly improves the model’s robustness.