A Deep Learning Approach to Mesh Segmentation

In the shape analysis community, decomposing a 3D into meaningful parts has become topic of interest. model segmentation is largely used in tasks such as deformation, partial matching, skeleton extraction, correspondence, annotation and texture mapping. Numerous approaches have attempted to provide better solutions; however, majority previous techniques handcrafted features, which are usually focused on particular attribute objects so difficult generalize. this paper, we propose three-stage approach for using Multi-view recurrent neural network automatically segment visually sub-meshes. The first stage involves normalizing scaling fit within unit sphere rendering object different views. Contrasting viewpoints, other hand, might not been associated, region could correlate totally distinct outcomes depending viewpoint. To address this, ran each view through (shared weights) CNN Bolster block order create probability boundary map. simulates area relationships between views, helps improve refine data. two, feature maps generated step correlated Recurrent Neural obtain compatible fine detail responses view. Finally, layer that fully connected return coherent edges, then back project produce final segmentation. Experiments Princeton Segmentation Benchmark dataset show our proposed method effective mesh tasks.