三维点云深度模型压缩算法

Compression method for three-dimensional point cloud deep model

随着计算机三维视觉的广泛应用,近几年基于深度学习的点云处理算法得到了大量研究,而耗时耗存储的缺陷较大程度限制了其在移动端的部署应用。基于改进损失函数的总体思路,提出了一种新的点云深度模型压缩框架,将知识蒸馏方法引入二值量化模型中,同时考虑点云聚合操作的特殊性引入了辅助损失项,改进的损失函数共包括预测损失项、蒸馏损失项和辅助损失项三部分。实验结果表明,和已有算法相比,所提算法可以获取更高的精度,同时对当前点云主流深度网络模型也具有良好的扩展性。

With the widespread application of computer three-dimensional vision,point cloud processing algorithms based on deep learning have attracted a lot of research in recent years,however,the time and storage consuming characteristics have greatly restricted its deployment and application on the mobile terminal devices.Based on the general idea of improving the loss function,a new point cloud deep model compression framework was proposed,and the knowledge distillation method was introduced into the binary quantization model.At the same time,considering the speciality of the point cloud aggregation operation,an auxiliary loss item was introduced.The improved loss function includes three parts:prediction loss,distillation loss and auxiliary loss.The experimental results show that,compared with the existing algorithms,the proposed algorithm can obtain higher accuracy,meanwhile,the application to current mainstream point cloud deep network models can also achieve good scalability.