LiDAR-PTQ: Post-Training Quantization for Point Cloud 3D Object Detection This paper proposes LiDAR-PTQ, a post-training quantization (PTQ) method specifically designed for 3D lidar-based object detection. The method addresses the challenges of quantizing 3D lidar-based detectors, which differ significantly from 2D vision tasks. The main components of LiDAR-PTQ include a sparsity-based calibration method to initialize quantization parameters, a task-guided global positive loss (TGPL) to reduce the disparity between final predictions before and after quantization, and an adaptive rounding-to-nearest operation to minimize layerwise reconstruction error. Extensive experiments show that LiDAR-PTQ achieves state-of-the-art quantization performance on CenterPoint (both pillar-based and voxel-based) with an accuracy almost the same as the FP32 model while enjoying 3× inference speedup. Additionally, LiDAR-PTQ is cost-effective, being 30× faster than the quantization-aware training method. The method is effective for both SPConv-based and SPConv-free 3D detection models. The paper also discusses the challenges of quantizing 3D lidar-based detectors, including the sparsity of point clouds, larger arithmetic range, and imbalance between foreground instances and large redundant background areas. The proposed method is evaluated on multiple datasets, including Waymo and nuScenes, demonstrating its effectiveness and efficiency. The results show that LiDAR-PTQ achieves state-of-the-art performance on these datasets, outperforming other PTQ methods. The method is also applicable to fully sparse detectors, such as FSD and FSD++. The paper concludes that LiDAR-PTQ is a valuable quantization tool for current mainstream grid-based 3D detectors and can push the development of practical deployment of 3D detection models on edge devices.LiDAR-PTQ: Post-Training Quantization for Point Cloud 3D Object Detection This paper proposes LiDAR-PTQ, a post-training quantization (PTQ) method specifically designed for 3D lidar-based object detection. The method addresses the challenges of quantizing 3D lidar-based detectors, which differ significantly from 2D vision tasks. The main components of LiDAR-PTQ include a sparsity-based calibration method to initialize quantization parameters, a task-guided global positive loss (TGPL) to reduce the disparity between final predictions before and after quantization, and an adaptive rounding-to-nearest operation to minimize layerwise reconstruction error. Extensive experiments show that LiDAR-PTQ achieves state-of-the-art quantization performance on CenterPoint (both pillar-based and voxel-based) with an accuracy almost the same as the FP32 model while enjoying 3× inference speedup. Additionally, LiDAR-PTQ is cost-effective, being 30× faster than the quantization-aware training method. The method is effective for both SPConv-based and SPConv-free 3D detection models. The paper also discusses the challenges of quantizing 3D lidar-based detectors, including the sparsity of point clouds, larger arithmetic range, and imbalance between foreground instances and large redundant background areas. The proposed method is evaluated on multiple datasets, including Waymo and nuScenes, demonstrating its effectiveness and efficiency. The results show that LiDAR-PTQ achieves state-of-the-art performance on these datasets, outperforming other PTQ methods. The method is also applicable to fully sparse detectors, such as FSD and FSD++. The paper concludes that LiDAR-PTQ is a valuable quantization tool for current mainstream grid-based 3D detectors and can push the development of practical deployment of 3D detection models on edge devices.