Research map / Detection and segmentation / 3D detection and driving
SwiftPillars: High-Efficiency Pillar Encoder for Lidar-Based 3D Detection
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 SwiftPillars, a pillar-based LiDAR 3D detector with a Swift Pillar Encoder (dual point/channel attention with lightweight matmul ops replacing PFN transposes) and a Multi-scale Aggregation Decoder, trading slight accuracy for much faster inference.
Problem
Current high-accuracy LiDAR 3D detectors use complex architectures (transformers, 3D sparse convolutions) that are slow and hard to deploy for real-time autonomous driving, while fast pillar methods like PointPillars lose accuracy and have I/O-bound encoders.
Approach
Fuse PointPillars' Point Feature Network with a lightweight point-wise and channel-wise attention module implemented with pooling and matrix multiplication, removing matrix transposes that cause memory-bound overhead, and aggregate multi-scale pseudo-image features with BiFPN-style aggregation/decoder blocks.
Key results
- Achieves 61.3% NDS and 53.2% mAP on nuScenes test set, about 3.2% mAP above PointPillars.
- Runs at 13.3 ms (75 FPS) on NVIDIA Tesla T4 with TensorRT; on average 26.58% faster than PointPillars across platforms.
- On KITTI val, moderate mAP 1.74% higher than PointPillars; on DAIR-V2X-I val, 0.71% higher.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace PointPillars with SwiftPillars | 3D LiDAR detection · nuScenes test | mAP: 30.5 -> 53.2 | Helped |
| replace PointPillars with SwiftPillars | 3D LiDAR detection · KITTI validation | mAP: 63.24 -> 64.98 | Helped |
| replace PointPillars with SwiftPillars | 3D LiDAR detection · DAIR-V2X-I validation | mAP: 51.31 -> 52.02 | Helped |
| replace PFN with SPE (ablation) | CenterPoint-pillar · nuScenes validation | mAP: 48.72 -> 49.75 | Helped |
| replace PFN with SPE (ablation) | CenterPoint-pillar · nuScenes validation | FPS: 70 -> 81 | Helped |
| replace RPN with MAD (ablation) | CenterPoint-pillar · nuScenes validation | mAP: 48.72 -> 51.16 | Helped |
Surprising findings
- Removing matrix transposes in the PFN-style encoder yields up to 100.2% speedup on NVIDIA A2, while RTX3080 and MLU370 show near-zero gains on some settings.
Limitations
- Accuracy trails voxel- and transformer-based methods on nuScenes (e.g., 53.2 mAP vs 65+ mAP), attributed to information loss in pillar encoding.
- Speed gains vary considerably across GPU platforms due to differing memory-bound characteristics.
Experimental setup
Datasets: nuScenes KITTI DAIR-V2X-I
Benchmarks: nuScenes KITTI DAIR-V2X-I
Metrics: mAP NDS FPS inference time (ms) AP 3D (R40)
Baselines: PointPillars CenterPoint-pillar SA-Det3D PillarAcc 3DSSD SASA TANet FocalConv OcTr LargeKernel3D TransFusion-L SphereFormer
Builds on: PointPillars CenterPoint BiFPN OpenPCDet
Closest papers by meaning
- Fully Sparse 3D Object Detection (NeurIPS 2022 · similarity 0.89)
- PointPillars: Fast Encoders for Object Detection From Point Clouds (CVPR 2019 · similarity 0.89)
- Voxel or Pillar: Exploring Efficient Point Cloud Representation for 3D Object Detection (AAAI 2024 · similarity 0.88)
- BEV-MAE: Bird’s Eye View Masked Autoencoders for Point Cloud Pre-training in Autonomous Driving Scenarios (AAAI 2024 · similarity 0.87)
- LSK3DNet: Towards Effective and Efficient 3D Perception with Large Sparse Kernels (CVPR 2024 · similarity 0.86)
- SAFDNet: A Simple and Effective Network for Fully Sparse 3D Object Detection (CVPR 2024 · similarity 0.86)
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.