[CVPR 2025] GBlobs: Explicit Local Structure via Gaussian Blobs for Improved Cross-Domain LiDAR-based 3D Object Detection
This repository features GBlobs, a parameter-free method designed to explicitly capture the local structure within point clouds. GBlobs lead to substantial improvements in domain generalization and cross-domain performance.
Abstract
LiDAR-based 3D detectors need large datasets for training, yet they struggle to generalize to novel domains. Domain Generalization (DG) aims to mitigate this by training detectors that are invariant to such domain shifts. Current DG approaches exclusively rely on global geometric features (point cloud Cartesian coordinates) as input features. Over-reliance on these global geometric features can, however, cause 3D detectors to prioritize object location and absolute position, resulting in poor cross-domain performance. To mitigate this, we propose to exploit explicit local point cloud structure for DG, in particular by encoding point cloud neighborhoods with Gaussian blobs, GBlobs. Our proposed formulation is highly efficient and requires no additional parameters. Without any bells and whistles, simply by integrating GBlobs in existing detectors, we beat the current state-of-the-art in challenging single-source DG benchmarks by over 21 mAP (Waymo->KITTI), 13 mAP (KITTI->Waymo), and 12 mAP (nuScenes->KITTI), without sacrificing in-domain performance. Additionally, GBlobs demonstrate exceptional performance in multi-source DG, surpassing the current state-of-the-art by 17, 12, and 5 mAP on Waymo, KITTI, and ONCE, respectively.
[2025-10] GBlobs, achieved state-of-the-art results in the IROS RoboSense Challange Track #3: Cross-Sensor Placement 3D Object Detection, securing 1st place. Check out our implementation in the separate branch robosense25_track3 or read our technical report.
This project extends the OpenPCDet framework. To get started, you'll need to set up your environment and prepare the necessary datasets by following OpenPCDet's installation instructions and getting started guide.
You can easily access pre-trained checkpoints directly from 🤗 Hugging Face:
# Make sure git-lfs is installed (https://git-lfs.com)
git lfs install
git clone https://huggingface.co/dmalic/GBlobs ckptsNote: Waymo Dataset License Agreement prevents us from sharing the models trained on Waymo.
To replicate the results from our paper, use the following commands:
python train.py --cfg_file <path to config>python test.py --cfg_file <path to config> --ckpt <path to checkpoint>For multi-GPU setups, convenience scripts are available:
- Single-node: torch_train.sh and torch_test.sh
- Multi-node: slurm_train_v2.sh and slurm_test_mgpu.sh
Note: While the Waymo Dataset License Agreement prevents us from sharing the pretrained models, you'll find that training with the default configurations yields similar results.
The results presented in Table 2 of the paper can be reproduced using the following:
| Task | 3D AP Car/Ped/Cyc | Config | Ckpt |
|---|---|---|---|
| Waymo→KITTI | 78.75/65.98/58.72 | cfgs/waymo-kitti_models/voxel_rcnn_with_centerhead_gblobs.yaml |
- |
| nuScenes→KITTI | 53.98/33.22/25.68 | cfgs/nuscenes-kitti_models/voxel_rcnn_with_centerhead_gblobs.yaml |
../ckpts/nuscenes-kitti_models/voxel_rcnn_with_centerhead_gblobs/checkpoint_epoch_30.pth |
| Waymo→nuScenes | 20.08/9.13/5.10 | cfgs/waymo-nuscenes_models/voxel_rcnn_with_centerhead_gblobs.yaml |
- |
KITTI→Waymo results from Table 3 of the paper can be reproduced with:
To reproduce the multi-source domain generalization results presented in Table 4 of the paper, use the following:
| Target* | mAP | Config | Ckpt |
|---|---|---|---|
| KITTI | 53.92 | cfgs/mdt3d_models/centerpoint_gblobs_kitti.yaml |
- |
| ONCE | 37.77 | cfgs/mdt3d_models/centerpoint_gblobs_nuscenes.yaml |
- |
| nuScenes | 8.15 | cfgs/mdt3d_models/centerpoint_gblobs_once.yaml |
- |
| Waymo | 23.81 | cfgs/mdt3d_models/centerpoint_gblobs_waymo.yaml |
../ckpts/mdt3d_models/centerpoint_gblobs_waymo/checkpoint_epoch_30.pth |
* Target refers to the dataset that was excluded from the training and used solely for evaluating the model's performance
If you find our work useful, please consider citing us:
@inproceedings{cvpr2025gblobs,
title={{GBlobs: Explicit Local Structure via Gaussian Blobs for Improved Cross-Domain LiDAR-based 3D Object Detection}},
author={Du\v{s}an Mali\'c and Christian Fruhwirth-Reisinger and Samuel Schulter and Horst Possegger},
booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition},
year={2025}
}