BEVDiffLoc

BEVDiffLoc: End-to-End LiDAR Global Localization in BEV View based on Diffusion Model

Environment

• python 3.10

• pytorch 2.1.2

• cuda 12.1

source install.sh

Dataset

We support the Oxford Radar RobotCar and NCLT datasets right now.

The data of the Oxford and NCLT dataset should be organized as follows:

data_root
├── 2019-01-11-14-02-26-radar-oxford-10k
│   ├── velodyne_left
│   │   ├── xxx.bin
│   │   ├── xxx.bin
│   ├── gps
│   │   ├── gps.csv
│   │   ├── ins.csv
│   ├── velodyne_left.timestamps
│   ├── merge_bev (Data prepare)
│   │   ├── xxx.png
│   │   ├── xxx.png
│   ├── merge_bev.txt (Data prepare)
├── Oxford_pose_stats.txt
├── train_split.txt
├── valid_split.txt

Data prepare

• NCLT: We use NCLT Sample Python Scripts to preprocess velodyne_sync to speed up data reading. We provided within it nclt_precess.py.

• Oxford&NCLT: We use merge_nclt.py and merge_oxford.py to generate local scenes for data augmentation.

Run

Download the pretrained ViT model

We initialize BEVDiffLoc's feature learner with DINOv2.

Train

accelerate launch --num_processes 1 --mixed_precision fp16 train_bev.py

Test

python test_bev.py

Acknowledgement

We appreciate the code of DiffLoc and BEVPlace++ they shared.