Research map / Point clouds / Segmentation and detection
GBlobs: Explicit Local Structure via Gaussian Blobs for Improved Cross-Domain LiDAR-based 3D Object Detection
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.
Replaces global Cartesian coordinate inputs of LiDAR 3D detectors with parameter-free Gaussian blob encodings (mean-offset and covariance) of local voxel neighborhoods to improve cross-domain generalization.
Problem
LiDAR-based 3D object detectors rely on global point coordinates, which bias them toward source-domain sensor configurations and object absolute position, hurting domain generalization to unseen datasets.
Approach
Encode each local point neighborhood (e.g., voxel) as a Gaussian blob with its mean offset and covariance matrix, a permutation-invariant, parameter-free local descriptor used instead of global xyz inputs, so the detector learns shape rather than absolute position.
Key results
- Voxel R-CNN with GBlobs reaches 73.48/67.82 mAP on Waymo->KITTI versus 65.41/46.79 baseline and 65.47/46.67 for Li et al.
- Multi-source CenterPoint with GBlobs reaches 30.91 average cross-domain mAP vs 22.87 for MDT3D, gaining ~17, 12 and 5 mAP on Waymo, KITTI and ONCE.
- On nuScenes->KITTI, GBlobs achieves 49.48/37.62 mAP vs 42.17/25.15 for Li et al. but nuScenes (sparse, mostly N<=2 points per voxel) gains are limited.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace global inputs with GBlobs | Voxel R-CNN · Waymo→KITTI | mAP 3D: 46.79 -> 67.82 | Helped |
| Replace global inputs with GBlobs | Voxel R-CNN · nuScenes→KITTI | mAP 3D: 21.07 -> 37.62 | Helped |
| Replace global inputs with GBlobs | Voxel R-CNN · Waymo→nuScenes | mAP 3D: 9.36 -> 11.44 | No clear effect |
| Replace global inputs with GBlobs | CenterPoint · Multi-source DG: Waymo, KITTI, ONCE | mAP: 22.87 -> 30.91 | Helped |
| Add local mean feature to global coordinates (ablation) | Voxel R-CNN · nuScenes→KITTI | mAP 3D: 21.07 -> 22.45 | Helped |
| Add covariance feature to global coordinates (ablation) | Voxel R-CNN · nuScenes→KITTI | mAP 3D: 21.07 -> 37.60 | Helped |
| Use local mean instead of global coordinates (ablation) | Voxel R-CNN · nuScenes→KITTI | mAP 3D: 21.07 -> 35.95 | Helped |
| Use covariance instead of global coordinates (ablation) | Voxel R-CNN · nuScenes→KITTI | mAP 3D: 21.07 -> 37.60 | Helped |
| Combine local mean and covariance, excluding global coordinates (ablation) | Voxel R-CNN · nuScenes→KITTI | mAP 3D: 21.07 -> 37.62 | Helped |
| Use GBlobs instead of relative-distance features (ablation) | Voxel R-CNN · Waymo→KITTI | mAP 3D: 65.34 -> 67.82 | Helped |
| Use GBlobs instead of surface-normal features (ablation) | Voxel R-CNN · Waymo→KITTI | mAP 3D: 63.96 -> 67.82 | Helped |
| Use GBlobs instead of RepSurf (ablation) | Voxel R-CNN · Waymo→KITTI | mAP 3D: 66.02 -> 67.82 | Helped |
Surprising findings
- Global xyz input features hurt cross-domain generalization; hybrid global+local features retain these drawbacks.
- In extreme sparse-to-sparser settings, global coordinates outperform local GBlobs features.
Limitations
- Degenerates to mean-only encoding when voxels contain fewer than 3 points, so performance is near baseline on sparse nuScenes.
- Underperforms when target data is much sparser than source (sparse-to-sparser), where global coordinates are preferred.
Experimental setup
Datasets: KITTI Waymo nuScenes ONCE
Benchmarks: KITTI Waymo nuScenes ONCE
Metrics: AP (BEV/3D) mAP
Baselines: Voxel R-CNN PA-DA 3D-VF Li et al. (multi-task DG) MDT3D CenterPoint RepSurf PointPillars SECOND Part-A2 DSVT
Builds on: Voxel R-CNN CenterPoint PointNet 3D-VF MDT3D
Closest papers by meaning
- Not All Points Are Equal: Learning Highly Efficient Point-Based Detectors for 3D LiDAR Point Clouds (CVPR 2022 · similarity 0.91)
- PillarNeXt: Rethinking Network Designs for 3D Object Detection in LiDAR Point Clouds (CVPR 2023 · similarity 0.89)
- An Empirical Study of the Generalization Ability of Lidar 3D Object Detectors to Unseen Domains (CVPR 2024 · similarity 0.88)
- Beyond Virtual Points: Depth-Enhanced LiDAR-only 3D Object Detection with Semi-Supervised Learning (Student Abstract) (AAAI 2025 · similarity 0.88)
- iDet3D: Towards Efficient Interactive Object Detection for LiDAR Point Clouds (AAAI 2024 · similarity 0.88)
- Pre-training LiDAR-based 3D Object Detectors through Colorization (ICLR 2024 · 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.