Research map / Architectures and efficiency / Attention and transformers
Towards Real-Time Full-Waveform LiDAR Transformers via Intensity-Guided Token Reduction and Physics-Aware Augmentation
Kazuma Ikeda, Kotaro Oishi, Ryosei Hara, Ryo Yoshida, Mariko Isogawa, Kazuma Ikeda
Read the paperCheck your idea against 137,217 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 a real-time Transformer for full-waveform LiDAR that uses intensity-guided token reduction and physics-aware augmentation (abstract only available).
Problem
Full-waveform LiDAR captures rich return signals robust to adverse conditions, but Transformer-based FWL models are computationally expensive for real-time use.
Approach
Reduces Transformer tokens using LiDAR return intensity as guidance and adds physics-aware data augmentation to FWL inputs; the abstract does not give further detail.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace prior SOTA (Ikeda et al.) with FWL-ToPM | FWL Transformer · Ghost-FWL | F1-mean and FPS: 0.503 -> 0.592 F1; 6.54 -> 53.2 FPS | Improved |
| Add intensity-based token pruning (rho_low=0.7) (ablation) | Ours w/o FWL-ToPM · Ghost-FWL | F1-mean: 0.555 -> 0.591 | Improved |
| Add sparse-aware token merging (ablation) | Ours w/o FWL-ToPM · Ghost-FWL | F1-mean: 0.555 -> 0.565 | Improved |
| Combine pruning and merging (ablation) | FWL-ToPM · Ghost-FWL | F1-mean and FPS: 20.3 FPS (pruning only) -> 53.2 FPS, F1 0.591 -> 0.592 | Improved |
| Use proposed pruning instead of DynamicViT (ablation) | Token pruning · Ghost-FWL | F1-mean: 0.480 -> 0.591 | Improved |
| Use proposed merging instead of ToMe (ablation) | Token merging · Ghost-FWL | F1-mean: 0.543 -> 0.565 | Improved |
| Increase pruning ratio to reach about 80 FPS (ablation) | Pruning only · Ghost-FWL | F1-mean | Hurt |
| Apply token pruning to a pretrained model without retraining (ablation) | FWL-ToPM · Ghost-FWL | F1-mean | Hurt |
| Apply token merging to a pretrained model without retraining (ablation) | FWL-ToPM · Ghost-FWL | F1-mean | No clear effect |
| Add FWLAug (CutMix + ShiftMix) (ablation) | FWL-ToPM · OOD test | F1-mean: 0.406 -> 0.452 | Improved |
| Add FWLAug (CutMix + ShiftMix) (ablation) | FWL-ToPM · OOD test Glass class | F1: 0.491 -> 0.575 | Improved |
| Add FWLCutMix only without ShiftMix (ablation) | FWL-ToPM · Ghost-FWL | F1-mean: 0.592 -> 0.586 | Hurt |
Closest papers by meaning
- TULIP: Transformer for Upsampling of LiDAR Point Clouds (CVPR 2024 · similarity 0.86)
- LiDAR-RT: Gaussian-based Ray Tracing for Dynamic LiDAR Re-simulation (CVPR 2025 · similarity 0.85)
- Neural Reconstruction of LiDAR Point Clouds under Jamming Attacks via Full-Waveform Representation and Simultaneous Laser Sensing (NeurIPS 2026 · similarity 0.85)
- LiT: Unifying LiDAR "Languages" with LiDAR Translator (NeurIPS 2024 · similarity 0.85)
- GS-LiDAR: Generating Realistic LiDAR Point Clouds with Panoramic Gaussian Splatting (ICLR 2025 · similarity 0.84)
- Efficient LiDAR Reflectance Compression via Scanning Serialization (ICML 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.
