[Paper Review] LiDAR-PTQ: Post-Training Quantization for Point Cloud 3D Object Detection

Sifan Zhou, Liang Li+7 more|arXiv (Cornell University)|Jan 29, 2024

3D Surveying and Cultural Heritage4 citations

TL;DR

This paper proposes LiDAR-PTQ, a post-training quantization method tailored for 3D LiDAR-based object detection that addresses performance degradation in INT8 quantization. By integrating sparsity-based calibration, Task-guided Global Positive Loss (TGPL), and adaptive rounding, it achieves INT8 model accuracy nearly matching FP32 with 3× inference speedup and 30× faster quantization than QAT, setting a new SOTA for PTQ in 3D detection.

ABSTRACT

Due to highly constrained computing power and memory, deploying 3D lidar-based detectors on edge devices equipped in autonomous vehicles and robots poses a crucial challenge. Being a convenient and straightforward model compression approach, Post-Training Quantization (PTQ) has been widely adopted in 2D vision tasks. However, applying it directly to 3D lidar-based tasks inevitably leads to performance degradation. As a remedy, we propose an effective PTQ method called LiDAR-PTQ, which is particularly curated for 3D lidar detection (both SPConv-based and SPConv-free). Our LiDAR-PTQ features three main components, extbf{(1)} a sparsity-based calibration method to determine the initialization of quantization parameters, extbf{(2)} a Task-guided Global Positive Loss (TGPL) to reduce the disparity between the final predictions before and after quantization, extbf{(3)} an adaptive rounding-to-nearest operation to minimize the layerwise reconstruction error. Extensive experiments demonstrate that our LiDAR-PTQ can achieve state-of-the-art quantization performance when applied to CenterPoint (both Pillar-based and Voxel-based). To our knowledge, for the very first time in lidar-based 3D detection tasks, the PTQ INT8 model's accuracy is almost the same as the FP32 model while enjoying $3 imes$ inference speedup. Moreover, our LiDAR-PTQ is cost-effective being $30 imes$ faster than the quantization-aware training method. Code will be released at \url{https://github.com/StiphyJay/LiDAR-PTQ}.

Motivation & Objective

• To address the severe performance drop in existing PTQ methods when applied to sparse, irregular 3D LiDAR point clouds.
• To develop a post-training quantization framework that maintains high detection accuracy while enabling efficient edge deployment.
• To overcome challenges from sparsity, large activation ranges, and foreground-background imbalance in 3D LiDAR feature maps.
• To achieve SOTA quantization performance comparable to full-precision models with minimal accuracy loss.

Proposed method

• Introduces a sparsity-based calibration method to initialize quantization parameters by analyzing the sparsity pattern of 3D point cloud features.
• Proposes Task-guided Global Positive Loss (TGPL), a loss function that minimizes the disparity between full-precision and quantized model outputs during calibration.
• Employs an adaptive rounding-to-nearest operation that reduces layerwise reconstruction error by optimizing rounding values per layer.
• Applies the method to both SPConv-based and SPConv-free 3D detectors, demonstrating generalization across architectures.
• Uses a small, unlabeled calibration set to enable fast, data-efficient quantization without retraining.
• Optimizes quantization parameters end-to-end using TGPL to align quantized model predictions with the original model's output distribution.

Experimental results

Research questions

• RQ1 Why do standard PTQ methods fail when directly applied to 3D LiDAR-based object detection models?
• RQ2 How can quantization parameters be initialized effectively to preserve model performance in sparse 3D feature spaces?
• RQ3 What loss function can best minimize the prediction gap between full-precision and quantized models in 3D detection?
• RQ4 Can adaptive rounding improve layerwise reconstruction and overall detection accuracy in 3D LiDAR models?
• RQ5 To what extent can PTQ achieve SOTA performance comparable to FP32 models in 3D LiDAR detection?

Key findings

• LiDAR-PTQ achieves INT8 quantization accuracy that is nearly identical to the full-precision FP32 model on CenterPoint (both pillar-based and voxel-based) across multiple datasets.
• The method delivers a 3× inference speedup with no significant accuracy drop, enabling real-time deployment on edge devices.
• Quantization via LiDAR-PTQ is 30× faster than quantization-aware training (QAT), making it highly efficient for industrial deployment.
• The sparsity-based calibration method effectively initializes quantization parameters by leveraging the inherent sparsity of 3D LiDAR features.
• TGPL significantly reduces the prediction disparity between full-precision and quantized models, improving detection mAP after quantization.
• LiDAR-PTQ generalizes well across both SPConv-based and SPConv-free 3D detection architectures, demonstrating broad applicability.