Research map / Point clouds / Point cloud classification
TULIP: Transformer for Upsampling of LiDAR Point Clouds
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 TULIP, a Swin-UNet-based range-image network for vertical LiDAR point cloud upsampling, using 1x4 row patches, rectangular windows, and circular padding.
Problem
LiDAR point cloud upsampling from low-resolution to high-resolution scans, where existing range-image super-resolution methods blur edges and produce invalid floating points.
Approach
Adapts Swin-UNet to range images with row-based (1x4) patch partition, non-square rectangular windows favoring the horizontal axis, circular padding in azimuth, and patch unmerging/pixel shuffle reconstruction, trained with pixel-wise L1 loss.
Key results
- On KITTI 4x upsampling, TULIP reaches MAE 0.4185, IoU 0.4174, CD 0.1207, versus ILN IoU 0.3342 and LiDAR-SR MAE 0.5674.
- On DurLAR, TULIP reaches IoU 0.3562 and CD 0.06484, outperforming ILN's CD of 0.0893; TULIP-L reaches the best CD of 0.06346.
- Row patches (1x4) cut MAE by 32.6%, improved IoU by 12.8%, and reduced CD by 28.3% relative to 4x4 patches in ablation on KITTI.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use 1 × 4 row patches instead of 4 × 4 patches (ablation) | TULIP · KITTI 4x | MAE: 0.7138 -> 0.4814 | Helped |
| Use rectangular instead of square attention windows (ablation) | TULIP · KITTI 4x | MAE: 0.4814 -> 0.4248 | Helped |
| Add circular padding, patch unmerging, and pixel shuffle (ablation) | TULIP · KITTI 4x | IoU: 0.4040 -> 0.4084 | Helped |
| Add Monte Carlo Dropout post-processing (ablation) | TULIP · KITTI 4x | MAE: 0.4227 -> 0.4185 | Helped |
| Increase model capacity to four Swin-Transformer layers in encoder and decoder (ablation) | TULIP · KITTI 4x | MAE: 0.4185 -> 0.3708 | Helped |
| Use ViT blocks instead of Swin Transformer blocks (ablation) | TULIP · KITTI 4x | MAE: 0.4185 -> 1.4484 | Hurt |
| Use TULIP instead of the strongest baseline for MAE | CARLA 4x · CARLA 4x | MAE: 0.8064 -> 0.7699 | Helped |
| Use TULIP instead of the strongest baseline for IoU | CARLA 4x · CARLA 4x | IoU: 0.5006 -> 0.5152 | Helped |
| Use TULIP instead of the strongest baseline for Chamfer Distance | CARLA 4x · CARLA 4x | Chamfer Distance: 0.1855 -> 0.1028 | Helped |
| Use TULIP instead of the strongest baseline for MAE | KITTI 4x · KITTI 4x | MAE: 0.5674 -> 0.4185 | Helped |
| Use TULIP instead of the strongest baseline for IoU | KITTI 4x · KITTI 4x | IoU: 0.3342 -> 0.4174 | Helped |
| Use TULIP instead of the strongest baseline for Chamfer Distance | KITTI 4x · KITTI 4x | Chamfer Distance: 0.1916 -> 0.1207 | Helped |
Surprising findings
- Image super-resolution networks like SwinIR perform worse than simple interpolation baselines on some range-image metrics, e.g., SwinIR CD 0.7347 on KITTI vs bilinear 0.6000 in CD.
- LIIF achieves lowest MAE among baselines on CARLA yet much worse 3D metrics than LiDAR-specific methods.
Limitations
- Performance degrades at ranges above 30m where LiDAR data is extremely sparse.
- Failure cases show noisy reconstruction and missed objects in irregular scenes.
- Only vertical upsampling is performed (factor 4 at fixed horizontal resolution).
Experimental setup
Datasets: KITTI DurLAR CARLA
Benchmarks: KITTI DurLAR CARLA
Metrics: MAE IoU Chamfer Distance
Baselines: Bilinear SRNO HAT SwinIR LIIF LiDAR-SR ILN ViT-UNet
Builds on: Swin-UNet Swin Transformer LiDAR-SR Implicit LiDAR Network
Closest papers by meaning
- TransLO: A Window-Based Masked Point Transformer Framework for Large-Scale LiDAR Odometry (AAAI 2023 · similarity 0.87)
- LiDAR-Net: A Real-scanned 3D Point Cloud Dataset for Indoor Scenes (CVPR 2024 · similarity 0.86)
- PLATYPUS: Progressive Local Surface Estimator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2025 · similarity 0.86)
- Pre-training a Density-Aware Pose Transformer for Robust LiDAR-based 3D Human Pose Estimation (AAAI 2025 · similarity 0.86)
- LiDAR-PTQ: Post-Training Quantization for Point Cloud 3D Object Detection (ICLR 2024 · similarity 0.85)
- Point Transformer V3: Simpler Faster Stronger (CVPR 2024 · similarity 0.85)
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.