Research map / Point clouds / Segmentation and detection
Using a Waffle Iron for Automotive Point Cloud Semantic Segmentation
Gilles Puy, Alexandre Boulch, Renaud Marlet
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add instance cutmix and polarmix augmentations (ablation) | WaffleIron-48-256 · SemanticKITTI val | mIoU: 62.5 -> 66.8 | Improved |
| add polarmix on top of instance cutmix (ablation) | WaffleIron-48-256 · SemanticKITTI val | mIoU: 65.3 -> 66.8 | Improved |
| use 5-dim input vectors instead of 3-dim (ablation) | WaffleIron-48-256 · SemanticKITTI val | mIoU: 66.8 -> 67.6 | Improved |
| add stochastic depth (ablation) | WaffleIron-48-256 · SemanticKITTI val | mIoU: 67.6 -> 68.0 | Improved |
| use 5-dim input, F=384 and stochastic depth (ablation) | WaffleIron-48 · nuScenes val | mIoU: 76.1 -> 77.6 | Improved |
| increase width and depth (ablation) | WaffleIron · nuScenes val | mIoU: 73.2 (L=6) -> 76.1 (L=48), F=256 | Improved |
| increase depth from 12 to 48 layers (ablation) | WaffleIron-F256 · SemanticKITTI val | mIoU: 62.6 -> 62.5 | No clear effect |
| vary 2D grid resolution rho within 40-80 cm (nuScenes) (ablation) | WaffleIron · nuScenes val | mIoU: varies by at most one point | No clear effect |
| compute three projections in parallel instead of in series (ablation) | WaffleIron-12-256 · nuScenes val | mIoU: 75.2 -> 73.4 | Hurt |
| reverse order of projections (ablation) | WaffleIron-12-256 · SemanticKITTI val | mIoU: 62.6 -> 60.9 | Hurt |
| project only in bird's-eye view at all layers (ablation) | WaffleIron-12-256 · nuScenes val | mIoU: 75.2 -> 74.8 | No clear effect |
| project only in BEV when strong augmentations are used (ablation) | WaffleIron-48-256 · SemanticKITTI val | mIoU: 66.8 -> 66.8 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- RPVNet — RPVNet: A Deep and Efficient Range-Point-Voxel Fusion Network for LiDAR Point Cloud Segmentation (ICCV 2021) · better on 1 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Efficient Outdoor 3D Point Cloud Semantic Segmentation for Critical Road Objects and Distributed Contexts (ECCV 2020 · similarity 0.89)
- PointGS: Semantic-Consistent Unsupervised 3D Point Cloud Segmentation with 3D Gaussian Splatting (CVPR 2026 · similarity 0.89)
- 3D Semantic Segmentation With Submanifold Sparse Convolutional Networks (CVPR 2018 · similarity 0.89)
- Domain Generalization of 3D Semantic Segmentation in Autonomous Driving (ICCV 2023 · similarity 0.88)
- SpSequenceNet: Semantic Segmentation Network on 4D Point Clouds (CVPR 2020 · similarity 0.88)
- 3D Annotation-Free Learning by Distilling 2D Open-Vocabulary Segmentation Models for Autonomous Driving (AAAI 2025 · 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.
