Research map / Point clouds / Point cloud classification
GLiDR: Topologically Regularized Graph Generative Network for Sparse 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 GLiDR, a graph generative network regularized by 0-dimensional persistent homology constraints that augments static LiDAR points along a connected-component backbone, recovering detail from 32x sparser scans.
Problem
Sparse LiDAR scans lose static structure detail and point density, hurting SLAM and navigation; existing generative and completion methods fail to preserve global LiDAR topology.
Approach
Represent a LiDAR range-scan as a k-NN graph with stacked graph layers for global receptivity, and add a 0-dimensional persistent homology loss on the output and intermediate embeddings so that augmented points form a single connected backbone along static structures.
Key results
- On CARLA-64 sparse scans, GLiDR reaches CD 0.91 versus 7.96 for the best baseline C-Net, and on KITTI sparse CD 1.60 versus 2.82 for C-Net.
- Ablation on CARLA: graph layers plus topology constraint reduce sparse error from 1.18 to 1.04 (Table 2), and replacing graph layers with ConvNets raises sparse error from 1.18 to 16.91.
- Inference runs at 4 ms per scan on an NVIDIA A100 GPU, with 30 ms preprocessing on a Xeon Silver 4208 CPU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GLiDR instead of the strongest listed baseline | Static point augmentation · KITTI dense | CD: 7.10 -> 3.57 | Helped |
| Use GLiDR instead of the strongest listed baseline | Static point augmentation · KITTI sparse | CD: 2.82 -> 1.60 | Helped |
| Use GLiDR instead of the strongest listed baseline | Static point augmentation · CARLA-64 dense | RMSE (10−3): 44.20 -> 29.23 | Helped |
| Use GLiDR instead of the strongest listed baseline | Static point augmentation · CARLA-64 sparse | CD: 15.67 -> 0.91 | Helped |
| Use GLiDR instead of the strongest listed baseline | Static point augmentation · ARD-16 dense | RMSE (10−3): 14.69 -> 18.80 | Hurt |
| Use GLiDR instead of the strongest listed baseline | Static point augmentation · ARD-16 sparse | CD: 0.33 -> 0.22 | Helped |
| Add the 0-dim PH constraint (ablation) | GLiDR graph network · CARLA-64 sparse | CD: 1.18 -> 1.04 | Helped |
| Add the 0-dim PH constraint (ablation) | GLiDR graph network · CARLA-64 dense | CD: 2.36 -> 2.08 | Helped |
| Replace graph layers with standard convolutional layers (ablation) | GLiDR · CARLA-64 sparse | reported ablation score (metric unspecified): 1.18 -> 16.91 | Hurt |
| Replace graph layers with standard convolutional layers (ablation) | GLiDR · CARLA-64 dense | reported ablation score (metric unspecified): 2.36 -> 11.3 | Hurt |
Surprising findings
- Replacing the graph layers with convolutional layers degrades sparse CARLA error by more than an order of magnitude (1.18 to 16.91), suggesting the graph component matters most.
Limitations
- Trained per dataset; the paper states it is unlikely to generalize across environments (outdoor, industrial, indoor) without retraining.
- Relies on semantic segmentation labels to build paired static/dynamic scans for training.
- Evaluation of SLAM and additional results are only in the supplementary material.
Experimental setup
Datasets: KITTI Odometry ARD-16 CARLA-64
Benchmarks: KITTI Odometry CARLA-64 ARD-16
Metrics: Chamfer Distance Earth Mover's Distance Jensen-Shannon Divergence RMSE Maximum Mean Discrepancy
Baselines: CP3 Topo-AE C-Net DSLR MOVES
Builds on: DSLR Topo-AE MOVES PointNet-style k-NN graph convolution
Closest papers by meaning
- GS-LiDAR: Generating Realistic LiDAR Point Clouds with Panoramic Gaussian Splatting (ICLR 2025 · similarity 0.85)
- DeepPointMap: Advancing LiDAR SLAM with Unified Neural Descriptors (AAAI 2024 · similarity 0.85)
- GeoNLF: Geometry guided Pose-Free Neural LiDAR Fields (NeurIPS 2024 · similarity 0.85)
- GBlobs: Explicit Local Structure via Gaussian Blobs for Improved Cross-Domain LiDAR-based 3D Object Detection (CVPR 2025 · similarity 0.84)
- Flexible Depth Completion for Sparse and Varying Point Densities (CVPR 2024 · similarity 0.84)
- GeoAuxNet: Towards Universal 3D Representation Learning for Multi-sensor Point Clouds (CVPR 2024 · 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.