Research map / Point clouds / Segmentation and detection
Spherical Frustum Sparse Convolution Network for LiDAR Point Cloud Semantic Segmentation
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 spherical frustum structure with hash-based storage, sparse convolution (SFC) and frustum farthest point sampling (F2PS) so 2D CNNs process LiDAR point clouds without quantized projection loss (SFCNet).
Problem
Spherical projection of LiDAR point clouds to 2D images drops multiple points landing on the same pixel, causing quantized information loss and poor segmentation of small objects.
Approach
Replace one-point-per-pixel spherical projection with a spherical frustum that keeps all points projected to the same 2D cell, stored via a hash table; apply sparse convolution that selects the nearest-range point in each neighboring frustum, and sample with stride over frustums plus FPS inside each frustum.
Key results
- SFCNet reaches 65.0% mIoU on SemanticKITTI test, above CENet (64.7%) and RangeViT (64.0%) among 2D projection methods.
- On nuScenes validation SFCNet achieves 75.9% mIoU, above RangeViT (75.2%) and SalsaNext (72.2%).
- Ablation on SemanticKITTI validation: baseline 56.2% mIoU, +SFC 60.5%, +F2PS 62.9%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace spherical projection with spherical frustum and use SFC (ablation) | Baseline on SemanticKITTI validation · SemanticKITTI validation | mIoU (%): 56.2 -> 60.5; increases 4.3% | Helped |
| replace stride-based 2D sampling with F2PS (ablation) | Baseline + SFC on SemanticKITTI validation · SemanticKITTI validation | mIoU (%): 60.5 -> 62.9; increases 2.4% | Helped |
| use SFCNet instead of RangeViT | 2D projection-based models · SemanticKITTI test | mIoU (%): 64.0 -> 65.0 | Helped |
| use SFCNet instead of CENet | 2D projection-based models · SemanticKITTI test | mIoU (%): 64.7 -> 65.0 | Helped |
| use SFCNet instead of RangeViT | 2D projection-based models · nuScenes validation | mIoU (%): 75.2 -> 75.9 | Helped |
| use SFCNet instead of SphereFormer | SemanticKITTI segmentation models · SemanticKITTI test | mIoU (%): 74.8 -> 65.0 | Hurt |
| use SFCNet instead of SphereFormer | nuScenes segmentation models · nuScenes validation | mIoU (%): 78.4 -> 75.9 | Hurt |
| use SFCNet instead of KNN-based post-processing | Baseline with quantized-loss restoration · SemanticKITTI validation | mIoU (%): 59.7 -> 62.9; 3.2% mIoU improvement | Helped |
| use SFCNet instead of KPConv refinement | Baseline with quantized-loss restoration · SemanticKITTI validation | mIoU (%): 60.1 -> 62.9; 2.8% mIoU improvement | Helped |
| limit spherical frustums to 2 or 4 points rather than preserving all points (ablation) | SFCNet · SemanticKITTI validation | mIoU (%): 61.0; 61.9 -> 62.9 | Hurt |
| use F2PS stride settings (2,1), (1,2), (2,4), or (4,2) instead of (2,2) (ablation) | SFCNet downsampling layers · SemanticKITTI validation | mIoU (%): 60.7; 61.3; 61.1; 62.4; 62.2; 62.3; 62.3; 62.6; 61.9; 60.5; 61.6; 61.9 -> 62.9 | Hurt |
| change the number of hash functions from 4 to 2, 3, or 5 (ablation) | SFCNet · SemanticKITTI validation | mIoU (%): 62.9; 62.9; 62.9; 62.9 | No clear effect |
Surprising findings
- Simply increasing projection resolution (to 2048 or 4096 width) lowers mIoU in the KNN post-processing baseline, so resolution expansion is not a fix for quantized loss.
Limitations
- Sparse convolution uses only the nearest point per neighboring frustum, which may limit receptive field and hurt wide-range classes like road and terrain.
- Evaluated only on supervised, unimodal LiDAR segmentation.
Experimental setup
Datasets: SemanticKITTI nuScenes
Benchmarks: SemanticKITTI nuScenes
Metrics: mIoU per-class IoU inference time (ms)
Baselines: RangeNet++ PolarNet SqueezeSegV3 SalsaNext KPRNet Lite-HDSeg RangeViT CENet Cylinder3D SphereFormer (AF)^2-S3Net PointNet++
Builds on: SqueezeSeg spherical projection RangeNet++ ResNet PointNet++ Lovász-Softmax loss
Closest papers by meaning
- Spherical Transformer for LiDAR-Based 3D Recognition (CVPR 2023 · similarity 0.88)
- Large-Scale Point Cloud Semantic Segmentation With Superpoint Graphs (CVPR 2018 · similarity 0.88)
- Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation (CVPR 2021 · similarity 0.88)
- Less Is More: Reducing Task and Model Complexity for 3D Point Cloud Semantic Segmentation (CVPR 2023 · similarity 0.87)
- Spherical Mask: Coarse-to-Fine 3D Point Cloud Instance Segmentation with Spherical Representation (CVPR 2024 · similarity 0.85)
- Taylor Series-Inspired Local Structure Fitting Network for Few-shot Point Cloud Semantic Segmentation (AAAI 2025 · similarity 0.84)
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.