Research map / Point clouds / Segmentation and detection

LiDAR-PTQ: Post-Training Quantization for Point Cloud 3D Object Detection

ICLR 2024 · Point clouds › Segmentation and detection · approach: Multimodal and open-vocabulary 3D detection · 4 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LiDAR-PTQ, a post-training INT8 quantization method for lidar 3D detectors (CenterPoint pillar/voxel, FSD) combining sparsity-aware max-min calibration with grid search, a task-guided global positive loss, and adaptive rounding, reaching near-FP32 mAPH on Waymo.

Problem

Post-training quantization of lidar-based 3D object detectors degrades severely because point cloud BEV features are highly sparse, have large arithmetic range, and foreground objects are small relative to the background; existing 2D PTQ methods transfer poorly.

Approach

Use max-min calibration (instead of entropy) with a lightweight grid search for scale initialization, then refine activation scales with a task-guided loss that distills the FP model's foreground-focused final predictions (Gaussian heatmap focal loss plus L1 box loss on pseudo-labels), and learn AdaRound-style adaptive rounding with a local layerwise reconstruction loss.

Key results

• On Waymo val, LiDAR-PTQ (8/8) reaches 60.12 mAPH/L2 for CenterPoint-Pillar vs 60.32 FP32, and 65.18 vs 65.25 for CenterPoint-Voxel, outperforming BRECQ, QDrop and PD-Quant.
• Entropy calibration drops CenterPoint-Pillar mAPH/L2 by 38.67 while max-min drops 7.41; LiDAR-PTQ closes the gap to within 0.2 mAPH.
• PTQ takes about 3 GPU-hours versus 94 GPU-hours for QAT on CenterPoint-Pillar, and INT8 gives about 3x inference speedup on Jetson AGX Orin (26.01 to 8.82 ms latency for Pillar).

Measured effects reported in the paper

Surprising findings

• Entropy calibration, the standard choice in 2D PTQ, performs much worse than simple max-min calibration on lidar detectors because zero-heavy sparse features make entropy clip informative large values.
• Quantization error of long-range objects (50m+) is much more severely hurt by entropy calibration (84.5% drop) than short-range objects.

Limitations

• Quantization still requires additional computation (grid search, reconstruction, TGPL fine-tuning) relative to simpler PTQ methods.
• Only INT8 is studied; low-bit quantization of 3D detectors is left to future work.
• Performance is bounded by the FP model, and the first and last layers are kept in full precision; the PFN layer of CenterPoint-Pillar is not quantized.

Experimental setup

Datasets: Waymo Open Dataset nuScenes SemanticKITTI

Benchmarks: Waymo Open Dataset val nuScenes val SemanticKITTI val

Metrics: mAP mAPH/L2 NDS mIoU latency FPS

Baselines: BRECQ QDrop PD-Quant QAT Entropy calibration Max-min calibration

Builds on: CenterPoint AdaRound BRECQ QDrop PD-Quant

Closest papers by meaning

• Point4Bit: Post Training 4-bit Quantization for Point Cloud 3D Detection (NeurIPS 2025 · similarity 0.89)
• Not All Points Are Equal: Learning Highly Efficient Point-Based Detectors for 3D LiDAR Point Clouds (CVPR 2022 · similarity 0.89)
• Point Density-Aware Voxels for LiDAR 3D Object Detection (CVPR 2022 · similarity 0.89)
• GBlobs: Explicit Local Structure via Gaussian Blobs for Improved Cross-Domain LiDAR-based 3D Object Detection (CVPR 2025 · similarity 0.87)
• Voxel or Pillar: Exploring Efficient Point Cloud Representation for 3D Object Detection (AAAI 2024 · similarity 0.87)
• Not All Tasks Quantize Equally: Fisher-Guided Quantization for Visual Geometry Transformer (NeurIPS 2026 · similarity 0.87)