一种优化的可拓展激光雷达点云可学习二值量化网络

Optimized Scalable and Learnable Binary Quantization Network for LiDAR Point Cloud

为解决激光雷达点云深度学习网络模型在移动端嵌入式设备部署存在的耗时耗存储问题,提出了一种激光雷达点云可学习二值量化网络模型。该模型基于特征的知识蒸馏,将全精度网络各层统计特征知识转移到二值量化网络,较大幅度地提升了量化精度;提出基于遗传算法的二值量化尺度因子恢复可学习优化算法,通过逐层搜索初始最优尺度恢复因子,并通过网络自学习大幅减少网络参数量;提出一种统计自适应池化损失最小化算法,包括量化网络自调节和全精度网络转移调节两种方式,以解决量化网络中池化信息损失较大的问题。实验结果表明,所提算法在获取高精度的同时实现了较大压缩比和加速比,可将PointNet大小压缩为原来的1/23、加速35倍以上,对其他点云主流深度网络具有良好的扩展性。

To solve the time-consuming and storage problems of the LiDAR point cloud deep learning network models in the deployment of embedded devices on the mobile terminal,a learnable binary quantization network model for LiDAR point clouds is proposed.The model refers to the idea of feature-based knowledge distillation and transfers the statistical feature knowledge of each layer of the full-precision network to the binary quantization network,which greatly improves quantification accuracy.A genetic-algorithm based learnable optimization algorithm for scale factor recovery of binary quantization is proposed,which searches for the initial optimal layer-wise scale recovery factor,and greatly reduces amount of network parameters through network self-learning.A statistical adaptive pooling loss minimization algorithm is proposed,including quantitative network self-adjustment and full-precision network transferring adjustment,which solves the problem of greater pooling information loss of quantitative networks.Experimental results show that the proposed algorithm achieves larger compression ratio and speedup ratio while obtaining high precision.Theoretically,it can compress PointNet by 23 times and accelerate it by 35 times at least or more,and also achieves good scalability for other mainstream point cloud deep networks.