Ziying Song · Caiyan Jia · Lin Liu · Shaoqing Xu · Lei Yang · Yadan Luo
GraphBEV++ is a unified multi-modal feature alignment framework for robust autonomous driving. It addresses the spatial and semantic misalignment between camera and LiDAR BEV features caused by calibration noise, inaccurate projection, and heterogeneous feature representations.
The framework combines two complementary components:
- LocalAlign-v2 corrects local correspondence errors during BEV construction through neighborhood-aware graph matching.
- GlobalAlign-v2 resolves accumulated representation-level inconsistencies during BEV fusion through deformable alignment or diffusion-based denoising.
GraphBEV++ supports both dense LSS-based and sparse query-based BEV representations, and generalizes across 3D detection, tracking, prediction, occupancy estimation, and end-to-end planning.
This repository provides two versions of GraphBEV++ for different autonomous driving settings:
- End-to-End Autonomous Driving Version: maintained in the
mainbranch, covering multi-task perception, prediction, occupancy estimation, and planning. - 3D Object Detection Version: maintained in the
3DODbranch, focusing on robust multi-modal 3D object detection under camera-LiDAR misalignment.
Please choose the corresponding branch according to your research task.
- Unified alignment: corrects misalignment at both BEV construction and feature fusion stages.
- Cross-paradigm design: supports LSS-based and query-based BEV representations.
- Noise robustness: retains strong detection performance under sensor calibration perturbations.
- Task generalization: improves perception, prediction, occupancy, and planning tasks.
- Practical efficiency: the LSS variant reaches 7.1 FPS on an NVIDIA A100 with limited alignment overhead.
| Component | Stage | Purpose | Variants |
|---|---|---|---|
| LocalAlign-v2 | BEV construction | Refines inaccurate projection and reference-point correspondence using neighborhood context | LSS, Query |
| GlobalAlign-v2 | BEV fusion | Aligns spatial shifts and semantic inconsistencies between heterogeneous BEV features | Deformable, Diffusion |
| Method | Clean mAP | Noisy mAP | Clean NDS | Noisy NDS | Relative mAP Drop |
|---|---|---|---|---|---|
| BEVFusion-MIT | 68.5 | 60.8 | 71.4 | 65.7 | 11.2% |
| SparseFusion | 70.4 | 64.7 | 72.8 | 67.1 | 8.1% |
| GraphBEV++ (LSS) | 70.7 | 69.3 | 73.2 | 72.3 | 2.0% |
| GraphBEV++ (Query) | 71.4 | 69.1 | 73.4 | 71.2 | 3.2% |
| Method | AMOTA | minADE | minFDE | MR | EPA |
|---|---|---|---|---|---|
| UniAD | 35.9 | 0.71 | 1.02 | 15.1 | 45.6 |
| FusionAD | 50.1 | 0.39 | 0.62 | 8.6 | 62.6 |
| GraphBEV++ (LSS) | 49.8 | 0.40 | 0.59 | 8.5 | 64.7 |
| GraphBEV++ (Query) | 51.1 | 0.38 | 0.52 | 7.7 | 64.5 |
The training and evaluation code is maintained in the official GraphBEVplus repository.
git clone https://github.com/adept-thu/GraphBEVplus.git
cd GraphBEVplusFollow the upstream guides to prepare the environment and nuScenes dataset:
Train the published query-based configuration on four GPUs:
CUDA_VISIBLE_DEVICES=0,1,2,3 \
bash tools/dist_train.sh projects/configs/query_graphbev/graphbev_query.py 4Evaluate a checkpoint:
CUDA_VISIBLE_DEVICES=0,1,2,3 \
bash tools/dist_test.sh projects/configs/query_graphbev/graphbev_query.py \
path/to/checkpoint.pth 4 --eval bboxThis repository also provides a dependency-free academic project page.
.
|-- index.html # Project page
|-- style.css # Page styles
|-- script.js # Tabs and citation-copy interaction
|-- resources/ # Method and result figures
|-- paper.pdf # Paper
`-- README.md
Open index.html directly for a local preview. To publish with GitHub Pages,
open the repository's Settings > Pages, select Deploy from a branch,
and choose the default branch with / (root).
The implementation builds on the OpenMMLab ecosystem and related BEV perception research, including BEVFusion, BEVFormer, and SparseFusion. Please also consider citing these works when using their code or configurations.
Please refer to the license and third-party notices in the official implementation.