3D point cloud is an important geometric data structure. In recent years, deep learning for 3D point cloud has attracted more and more attention and has been widely applied in autonomous driving, robotics and some other fields. Existing 3D point cloud deep learning models can guarantee a relatively high accuracy during training while sacrificing the inference speed due to the large amount of model parameters, which poses a challenge for real-time semantic segmentation of 3D point clouds. To accelerate the inference process, alternative solutions optimize the model through model pruning (reducing the amount of parameters and calculations of the model) or model quantization (compressing the amount of data size), which will inevitably lead to loss of accuracy. Therefore, a natural thought is to balance inference speed and training accuracy. In the paper, we propose a plug-and-play dynamic multi-branch neural network module, which can improve the training accuracy without pruning the size of the inference model or hampering the speed of inference. This module refers to the idea of structural re-parameterization, which we named it Rep-PointNet Module. In general, Rep-PointNet Module can efficiently improve the training accuracy of the model while performing lossless compression of the model. The experiment results show that our method beats the state-of-the-art on the ShapeNet dataset with rapid inference speed. The gains in the overall accuracy metric brought by the Rep-PointNet Module are about 1.5% and 2% on the classification task and the segmentation task, respectively.
'/%3e%3cuse xlink:href='%23a' transform='rotate(23.036 .093 25.536)'/%3e%3cuse xlink:href='%23a' transform='rotate(45.87 1.273 16.18)'/%3e%3cuse xlink:href='%23a' transform='rotate(69.945 .996 12.078)'/%3e%3cuse xlink:href='%23a' transform='rotate(20.66 -19.689 31.932)'/%3e%3c/svg%3e)
