Research map / Point clouds / Segmentation and detection
RangeViT: Towards Vision Transformers for 3D Semantic Segmentation in Autonomous Driving
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 a convolutional stem instead of linear patch embedding (ablation) | RangeViT on nuScenes · nuScenes validation | mIoU: 65.52 69.82 | Helped |
| Add the UpConv decoder instead of a linear decoder (ablation) | RangeViT on nuScenes · nuScenes validation | mIoU: 69.8 → 73.83 | Helped |
| Replace K-NN post-processing with the KPConv 3D refiner (ablation) | RangeViT on nuScenes · nuScenes validation | mIoU: 73.83 → 74.60 | Helped |
| Use a wider stem, decoder, and refiner (ablation) | RangeViT · nuScenes validation | mIoU: 74.00; 74.12; 74.60; 75.21 | Helped |
| Increase D_h beyond 128 (ablation) | RangeViT · SemanticKITTI validation | mIoU: 60.73; 60.45; 60.55 | Hurt |
| Use finer, rectangular 2 × 8 tokens rather than 16 × 16 tokens (ablation) | RangeViT · nuScenes validation | mIoU: 68.45 → 75.21 | Helped |
| Initialize the ViT with image pre-training rather than random weights (ablation) | RangeViT on nuScenes · nuScenes validation | mIoU: 72.37 → 74.77; 72.37 → 75.21 | Helped |
| Keep attention layers frozen and fine-tune FFN layers rather than fully fine-tuning the encoder (ablation) | RangeViT with Cityscapes pre-training · nuScenes validation | mIoU: 75.21 → 75.47 | Helped |
| Replace ViT-S with ResNet-50 (ablation) | RangeViT on nuScenes · nuScenes validation | mIoU: 74.77 → 72.30 | Hurt |
| Remove the transformer encoder (ablation) | RangeViT on nuScenes · nuScenes validation | mIoU: 67.88 → 53.73 | Hurt |
| Remove the classification token (ablation) | RangeViT with Cityscapes pre-training · nuScenes validation | mIoU: 75.21% → 74.64% | Hurt |
| Remove the classification token (ablation) | RangeViT with random initialization · nuScenes validation | mIoU: 72.37% → 72.24% | Hurt |
Closest papers by meaning
- MFINet: Multi-view Fusion and 2D–3D Interaction Enhancement for Real-Time LiDAR Semantic Segmentation (AAAI 2026 · similarity 0.89)
- Less Is More: Reducing Task and Model Complexity for 3D Point Cloud Semantic Segmentation (CVPR 2023 · similarity 0.89)
- Spiral: Semantic-Aware Progressive LiDAR Scene Generation and Understanding (NeurIPS 2025 · similarity 0.88)
- VLM2Scene: Self-Supervised Image-Text-LiDAR Learning with Foundation Models for Autonomous Driving Scene Understanding (AAAI 2024 · similarity 0.87)
- HiLoTs: High-Low Temporal Sensitive Representation Learning for Semi-Supervised LiDAR Segmentation in Autonomous Driving (CVPR 2025 · similarity 0.87)
- PC-BEV: An Efficient Polar-Cartesian BEV Fusion Framework for LiDAR Semantic Segmentation (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.