FFMeshSeg: A 3D feature fusion mesh segmentation model based on cross element attention

The 3D mesh segmentation is pivotal in diverse fields such as industrial design, intelligent manufacturing, and other technological domains. Many 3D modeling operations are contingent on efficient and precise mesh segmentation. However, the inherent complexity and irregularity of 3D mesh data render manual segmentation exceedingly costly and labor-intensive. Prevailing deep learning-based methods continue to face challenges, including suboptimal utilization of raw data, inefficiencies, and limited overall accuracy. To address these challenges, this paper introduces FFMeshSeg, an innovative 3D mesh segmentation model based on cross-element attention and feature fusion. The model incorporates three distinct feature aggregation branches, each designed to extract topological element features from grid vertices, edges, and faces. Subsequently, a cross-element feature fusion method grounded in a cross-attention mechanism is proposed. Additionally, a cross-element adjacency mask matrix is constructed to facilitate sparse attention feature fusion within local neighborhoods, resulting in more precise boundary segmentation. Experimental results demonstrate that the per-face segmentation accuracy of the FFMeshSeg in the COSEG dataset compared to SOTA models improves by at least 1% in the Vases and Chairs subsets. Furthermore, while maintaining a relative low amount of parameters, the computational efficiency has increased by over 50% in specific dataset compared to the classical methods.