Post-Training Quantization Based on Knowledge Distillation for Point Cloud 3D Object Detection

Deploying LiDAR-based detectors on edge devices presents significant challenges due to limited computing power and memory. Post-training quantization (PTQ) requires only a small dataset for calibration and does not require end-to-end retraining. As a result, it has been widely used as a lightweight and practical model compression technique in various 2D object detection. However, applying PTQ in LiDAR-based 3D object detection inevitably leads to decreased accuracy, especially in transformer-based 3D object detection. Therefore, we propose a universal PTQ for point cloud networks, called LKD-PTQ. LKDPTQ consists of the following three components, (1) point cloud sparsity-based calibration method Spercentile, for initializing the quantitative parameters, (2) a distillation loss tailored for 3D object detection networks to facilitate knowledge transfer between the student and teacher networks, (3) a knowledge calibration module to address knowledge transfer biases caused by the features of the detection head. The experiments demonstrate that the proposed LKD-PTQ achieves better accuracy compared with previous methods and can be effectively applied to both transformer-based and sparse convolution-based point cloud networks.