A light and efficient attention model for 3D shape part-segmentation

Part segmentation is one of the important tasks in 3D shape analysis. Prior works mostly rely on complex local modeling to learn features directly from raw mesh or point cloud. We proposed a simple, fast and robust approach for 3D shape part segmentation without sophisticated local geometric modeling and ingenious networks. Our main idea is to learn more discriminative features from different geometric descriptors by sparse convolution and attention mechanism. Specifically, an effective shape representation block, consists of a features embedding module, two attention module. The coordinates are fed into the features embedding module to produce embedding vectors. The attention head to learn the intrinsic relations between different descriptors and produce more informative feature maps. In the lower branch, We use some geometric descriptors as local features. And a classification head predicts the corresponding label. Conditional Random Field (CRF) is applied to optimize the segmentation of the network. The experiment results on Princeton Shape Benchmark(PSB) demonstrate that our architecture outperforms different methods, with higher accuracy, lower complexity and faster speed.