The paper "LiDAR-PTQ: Post-Training Quantization for Point Cloud 3D Object Detection" addresses the challenge of deploying 3D LiDAR-based detectors on edge devices due to limited computing power and memory. Post-Training Quantization (PTQ) is a widely adopted model compression approach in 2D vision tasks, but its direct application to 3D LiDAR-based tasks leads to performance degradation. To address this, the authors propose LiDAR-PTQ, a specialized PTQ method for 3D LiDAR detection, which includes three main components: 1. **Sparsity-based Calibration**: A method to determine the initialization of quantization parameters using a Max-min calibrator and lightweight grid search. 2. **Task-guided Global Positive Loss (TGPL)**: A function to optimize quantization parameters in the model space, minimizing the output disparity between the quantized and full-precision models. 3. **Adaptive Rounding-to-nearest**: An operation to mitigate layer-wise reconstruction error by adjusting the rounding value. Experiments on the Waymo dataset demonstrate that LiDAR-PTQ achieves state-of-the-art performance on CenterPoint (both Pillar-based and Voxel-based) models, with the PTQ INT8 model's accuracy nearly matching the FP32 model while providing a 3× inference speedup. LiDAR-PTQ is also 30× faster than quantization-aware training (QAT) methods, making it cost-effective. The method is applicable to both SPCov-based and SPCov-free 3D detection models and shows promising results on fully sparse detectors. The paper concludes by highlighting the effectiveness and efficiency of LiDAR-PTQ for practical deployment of 3D detection models on edge devices.The paper "LiDAR-PTQ: Post-Training Quantization for Point Cloud 3D Object Detection" addresses the challenge of deploying 3D LiDAR-based detectors on edge devices due to limited computing power and memory. Post-Training Quantization (PTQ) is a widely adopted model compression approach in 2D vision tasks, but its direct application to 3D LiDAR-based tasks leads to performance degradation. To address this, the authors propose LiDAR-PTQ, a specialized PTQ method for 3D LiDAR detection, which includes three main components: 1. **Sparsity-based Calibration**: A method to determine the initialization of quantization parameters using a Max-min calibrator and lightweight grid search. 2. **Task-guided Global Positive Loss (TGPL)**: A function to optimize quantization parameters in the model space, minimizing the output disparity between the quantized and full-precision models. 3. **Adaptive Rounding-to-nearest**: An operation to mitigate layer-wise reconstruction error by adjusting the rounding value. Experiments on the Waymo dataset demonstrate that LiDAR-PTQ achieves state-of-the-art performance on CenterPoint (both Pillar-based and Voxel-based) models, with the PTQ INT8 model's accuracy nearly matching the FP32 model while providing a 3× inference speedup. LiDAR-PTQ is also 30× faster than quantization-aware training (QAT) methods, making it cost-effective. The method is applicable to both SPCov-based and SPCov-free 3D detection models and shows promising results on fully sparse detectors. The paper concludes by highlighting the effectiveness and efficiency of LiDAR-PTQ for practical deployment of 3D detection models on edge devices.