Research map / Detection and segmentation / 3D detection and driving
PC-BEV: An Efficient Polar-Cartesian BEV Fusion Framework for LiDAR 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 PC-BEV, a BEV-only LiDAR segmentation framework that fuses Polar and Cartesian BEV branches via precomputed remap operations and a hybrid Transformer-CNN backbone, reaching SOTA-level mIoU on SemanticKITTI and nuScenes with faster inference than point-based multiview fusion.
Problem
Multiview LiDAR segmentation fusion (e.g., range view and BEV) relies on computationally expensive point-based grid sampling and scatter operations because views lack fixed correspondences, which hinders real-time deployment.
Approach
Fuse Polar and Cartesian partitionings within a single BEV space, exploiting their fixed grid-to-grid correspondence to precompute remap indices for dense, memory-friendly feature alignment instead of point-based interaction; add a Transformer-then-U-Net CNN hybrid backbone.
Key results
- PC-BEV reaches 78.8% mIoU on nuScenes validation, surpassing AMVNet (76.1%) and RPVNet (77.6%).
- PC-BEV achieves 66.4% mIoU on SemanticKITTI validation and 67.2% on test, versus 65.2% (AMVNet val) and 68.3% for CPGNet on test.
- Remap-based fusion takes 0.04 ms versus 6.8-8.0 ms for point-based grid sampling plus scatter, a claimed 170x speedup; full model latency is 31 ms vs 39 ms for point-based fusion.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fuse Polar and Cartesian BEV branches instead of using the Polar branch alone (ablation) | PC-BEV · SemanticKITTI validation | mIoU: 61.2 -> 66.4 | Helped |
| Fuse Polar and Cartesian BEV branches instead of using the Cartesian branch alone (ablation) | PC-BEV · SemanticKITTI validation | mIoU: 56.5 -> 66.4 | Helped |
| Fuse Polar and Cartesian partitioning strategies (ablation) | PC-BEV · nuScenes validation | mIoU: 75.0 -> 76.2 | Helped |
| Add remap-based feature fusion to the two-branch model (ablation) | PC-BEV · nuScenes validation | mIoU: 76.2 -> 78.2 | Helped |
| Add Transformer-CNN Mixture Architecture (ablation) | PC-BEV · nuScenes validation | mIoU: 78.2 -> 78.8 | Helped |
| Replace remap-based fusion with point-based fusion (ablation) | PC-BEV · nuScenes validation | latency: 29 -> 39 | Hurt |
| Use remap-based rather than point-based feature interaction | BEV feature interaction · BEV features, 0.2m voxel resolution | latency: 6.8 -> 0.04; 8.0 -> 0.04 | Helped |
| Use PC-BEV rather than AMVNet | LiDAR semantic segmentation · SemanticKITTI validation | mIoU: 65.2 -> 66.4 | Helped |
| Use PC-BEV rather than RPVNet | LiDAR semantic segmentation · nuScenes validation | mIoU: 77.6 -> 78.8 | Helped |
| Use PC-BEV rather than CPGNet | LiDAR semantic segmentation · SemanticKITTI test | mIoU: 68.3 -> 67.2 | Hurt |
Surprising findings
- On the SemanticKITTI test set, PC-BEV (67.2%) falls short of CPGNet (68.3%), which is the top entry in that comparison.
Limitations
- Performance on SemanticKITTI test trails CPGNet, the strongest multiview baseline there.
- Gains are reported on mIoU only; the speed comparison relies on a specific operator implementation and GPUs.
- Evaluated only on LiDAR datasets; extension to multi-camera BEV is left as future work.
Experimental setup
Datasets: SemanticKITTI nuScenes-lidarseg
Benchmarks: SemanticKITTI nuScenes-lidarseg
Metrics: mIoU latency (ms)
Baselines: RangeNet++ PolarNet GFNet AMVNet RPVNet CPGNet RangeViT CENet MaskRange KPRNet SalsaNext PCSCNet
Builds on: PolarNet AMVNet CPGNet GFNet SalsaNext
Closest papers by meaning
- PolarNet: An Improved Grid Representation for Online LiDAR Point Clouds Semantic Segmentation (CVPR 2020 · similarity 0.90)
- BEVDilation: LiDAR-Centric Multi-Modal Fusion for 3D Object Detection (AAAI 2026 · similarity 0.90)
- MFINet: Multi-view Fusion and 2D–3D Interaction Enhancement for Real-Time LiDAR Semantic Segmentation (AAAI 2026 · similarity 0.90)
- Panoptic-PolarNet: Proposal-Free LiDAR Point Cloud Panoptic Segmentation (CVPR 2021 · similarity 0.89)
- Adaptive-Smooth LiDAR-Camera Knowledge Distillation with Heterogeneous Fusion for Multi-View 3D Object Detection (AAAI 2026 · similarity 0.89)
- TASeg: Temporal Aggregation Network for LiDAR Semantic Segmentation (CVPR 2024 · similarity 0.87)
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.