Research map / Detection and segmentation / 3D detection and driving
PillarHist: A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram
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 PillarHist, a plug-in pillar feature encoder that replaces max-pooled point MLP with per-pillar height histograms (plus intensity histograms and pillar center), improving 3D detection accuracy and int8 quantization robustness on KITTI, nuScenes and Waymo.
Problem
Pillar-based LiDAR 3D detectors lose height-dimension information through element-wise max-pooling in pillar feature encoding and have large numerical range differences among PFE inputs, which hurts accuracy and quantization.
Approach
Encode each pillar by a height-binned point-count histogram and an intensity-weighted histogram, concatenated with pillar center coordinates and projected once per pillar by a linear layer, removing point-wise MLP and max-pooling while bounding the input value range for quantization.
Key results
- On nuScenes val, PillarHist raises NDS by 1.7 (PointPillars), 1.4 (CP-Pillar) and 1.6 (PillarNet); under 8-bit PTQ the drop shrinks from 7.3 to 1.6 NDS for PointPillars.
- On KITTI val, PillarHist reaches 81.42 car 3D AP (easy/mod/hard 88.80/79.13/76.30 R40 mAP 81.42) at 0.065 GFLOPs versus 0.152 for the original PFE.
- On Waymo val (20% training data), CenterPoint-Pillar with PillarHist improves mAP L2 by 1.2 and mAPH L2 by 1.5.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the PointPillars PFE with PillarHist | PointPillars · nuScenes test | NDS: 45.3 -> 47.0 | Helped |
| Replace the CenterPoint PFE with PillarHist | CenterPoint · nuScenes test | NDS: 65.5 -> 66.6 | Helped |
| Replace the PillarNet-34 PFE with PillarHist | PillarNet-34 · nuScenes test | NDS: 71.4 -> 72.0 | Helped |
| Replace PFE with PillarHist and quantize weights and activations to 8-bit | PointPillars · nuScenes val | NDS: 52.1 -> 44.8; 53.8 -> 52.2 | Mixed |
| Replace PFE with PillarHist and quantize weights and activations to 8-bit | CP-Pillar · nuScenes val | NDS: 56.4 -> 48.6; 57.8 -> 56.3 | Mixed |
| Replace PFE with PillarHist and quantize weights and activations to 8-bit | PillarNet · nuScenes val | NDS: 61.4 -> 54.6; 63.0 -> 61.6 | Mixed |
| Replace CP-Pillar PFE with PillarHist | CP-Pillar · Waymo val, 20% training data | mAPH (L2), Mean: 55.9 -> 57.3 | Helped |
| Add intensity histogram input (ablation) | PillarHist · KITTI val, Pedestrian 3DR40 | mAP: 49.62 -> 51.07 | Helped |
| Add intensity histogram input (ablation) | PillarHist · KITTI val, Car 3DR40 | mAP: 80.81 -> 81.41 | Helped |
| Add pillar center coordinates (ablation) | PillarHist · KITTI val, Pedestrian 3DR40 | mAP: 49.62 -> 51.07 | Helped |
| Add pillar center coordinates (ablation) | PillarHist · KITTI val, Car 3DR40 | mAP: 80.81 -> 81.41 | Helped |
Surprising findings
- Using only the z (height) coordinate inputs to PFE achieves performance comparable to the full baseline, while x- or y-only inputs cause large drops.
Limitations
- Gains are modest, averaging about 1 point across baselines.
- Quantization evaluated only with simple post-training RTN int8 on nuScenes val; no advanced quantization methods explored.
Experimental setup
Datasets: KITTI nuScenes Waymo
Benchmarks: KITTI val nuScenes test nuScenes val Waymo val
Metrics: AP3D R40 AP3D R11 NDS mAP latency GFLOPs mAPH
Baselines: PointPillars CenterPoint-Pillar PillarNet FastPillars TANet TA module PiFeNet PAA module FG-PFE VoxelNeXt Focals Conv
Builds on: PointPillars CenterPoint PillarNet FastPillars OpenPCDet
Closest papers by meaning
- SwiftPillars: High-Efficiency Pillar Encoder for Lidar-Based 3D Detection (AAAI 2024 · similarity 0.86)
- Voxel or Pillar: Exploring Efficient Point Cloud Representation for 3D Object Detection (AAAI 2024 · similarity 0.86)
- PillarNeXt: Rethinking Network Designs for 3D Object Detection in LiDAR Point Clouds (CVPR 2023 · similarity 0.85)
- LiDAR-PTQ: Post-Training Quantization for Point Cloud 3D Object Detection (ICLR 2024 · similarity 0.83)
- DALDet: Depth-Aware Learning Based Object Detection for Autonomous Driving (AAAI 2024 · similarity 0.83)
- BEVSpread: Spread Voxel Pooling for Bird's-Eye-View Representation in Vision-based Roadside 3D Object Detection (CVPR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.