Optimized Scalable and Learnable Binary Quantization Network for LiDAR Point Cloud

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 layerwise 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.