Code Release of "XLD: A Cross-Lane Dataset for Benchmarking Novel Driving View Synthesis"

Hao Li, Ming Yuan, Yan Zhang, Chenming Wu, Chen Zhao, Chunyu Song, Haocheng Feng, Errui Ding, Dingwen Zhang, Jingdong Wang

Project page | Paper | Data

This repository contains the official authors data preprocess tools for NeRFStudio and Gaussian Splatting methods.

Data Preprocess

1. Unzip the compressed files (e.g. carla_pic_0603_Town**.zip).
2. Download intrinscis.zip and extrinsics.zip from data link and unzip them.
3. Then move intrinscis and extrinsics to the corresponding folder.

cp -r intrinsics data/carla_pic_0603_Town01
cp -r extrinsics data/carla_pic_0603_Town01
cp -r intrinsics data/carla_pic_0603_Town02
cp -r extrinsics data/carla_pic_0603_Town02
cp -r intrinsics data/carla_pic_0603_Town03
cp -r extrinsics data/carla_pic_0603_Town03
cp -r intrinsics data/carla_pic_0603_Town04
cp -r extrinsics data/carla_pic_0603_Town04
cp -r intrinsics data/carla_pic_0603_Town05
cp -r extrinsics data/carla_pic_0603_Town05
cp -r intrinsics data/carla_pic_0603_Town10
cp -r extrinsics data/carla_pic_0603_Town10

The data tree should look like this:

generate_offset_0_test.py
rename.py
convert_XLD_transforms.py
data
├── carlc_pic_0603_Town0N
├────────────|─────────── intrinsics
├────────────|─────────────── | ───────── 0.txt
├────────────|─────────────── | ───────── 1.txt
├────────────|─────────────── | ───────── 2.txt
├────────────|─────────── extrinsics
├────────────|─────────────── | ───────── 0.txt
├────────────|─────────────── | ───────── 1.txt
├────────────|─────────────── | ───────── 2.txt
├────────────|─────────── train_pic
├────────────|─────────────── | ───────── camera_extrinscs_00000N.json
├────────────|─────────────── | ───────── cameraX_0000N.png
├────────────|─────────────── | ───────── lidar_0000N.ply
├────────────|─────────── test_pic
├────────────|─────────────── | ───────── offset_left_0m (don't exist, run generate_offset_0_test.py)
├────────────|─────────────── | ─────────────── | ───────── camera_extrinscs_00000N.json
├────────────|─────────────── | ─────────────── | ───────── cameraX_0000N.png
├────────────|─────────────── | ─────────────── | ───────── lidar_0000N.ply
├────────────|─────────────── | ───────── offset_left_1m
├────────────|─────────────── | ─────────────── | ───────── camera_extrinscs_00000N.json
├────────────|─────────────── | ─────────────── | ───────── cameraX_0000N.png
├────────────|─────────────── | ─────────────── | ───────── lidar_0000N.ply
├────────────|─────────────── | ───────── offset_left_4m
├────────────|─────────────── | ─────────────── | ───────── camera_extrinscs_00000N.json
├────────────|─────────────── | ─────────────── | ───────── cameraX_0000N.png
├────────────|─────────────── | ─────────────── | ───────── lidar_0000N.ply

Generate test set in 0m offset.

python generate_offset_0_test.py

Rename train images to train_** and test images to eval_** as the format of NeRFStudio.

python rename.py

Generate transforms.json for NeRF and 3D-GS.

python convert_XLD_transforms.py

Notably, change the arguments num_cams and camera_list for different camera settings, offset_meters for different offset settings.

###### multi cameras, offsets=1m     ######
generate_json_file(scene_path = 'data/carla_pic_0603_Town10', num_cams=3, camera_list=[1,0,2], offset_meters=1)
###### front-only camera, offsets=1m ######
generate_json_file(scene_path = 'data/carla_pic_0603_Town10', num_cams=1, camera_list=[0], offset_meters=1)

Model Training

UC-NeRF training

We provide a modified version of UC-NeRF to train on the XLD dataset flexibly.

Please refer to the readme.md of the modified UC-NeRF on this page.

EmerNeRF training

We provide a modified version of EmerNeRF to train on the XLD dataset flexibly.

Please refer to the readme.md of the modified EmerNeRF on this page.

NeRFacto training

cd carla_pic_0603_Town01_cam1
#### train #####
CUDA_VISIBLE_DEVICES=1 ns-train nerfacto --pipeline.model.camera-optimizer.mode off --experiment-name nerfacto_offset_1m nerfstudio-data --eval-mode filename  --data ./
#### eval #####
CUDA_VISIBLE_DEVICES=2 ns-eval --load-config outputs/nerfacto_offset_1m/nerfacto/2024-06-06_145158/config.yml    --output-path ./nerfacto_offset_1/result.json --render-output-path ./nerfacto_offset_1

Instant-NGP training

cd carla_pic_0603_Town01_cam1
#### train #####
CUDA_VISIBLE_DEVICES=0 ns-train instant-ngp --experiment-name instant_ngp_offset_1 nerfstudio-data --eval-mode filename  --data ./
#### eval #####
CUDA_VISIBLE_DEVICES=0 ns-eval --load-config outputs/instant_ngp_offset_1/instant-ngp/2024-06-06_165753/config.yml     --output-path ./instant-ngp_offset_1m/result.json --render-output-path ./instant-ngp_offset_1m

Gaussian methods training

Colmap-format data proprocessing

Run python scripts to transforms.json to 3D-GS format.

python convert_XLD_transforms.py.  # generate transforms.json
python merge_lidar.py              # merge 3D LiDAR point cloud as initialization 
python transform2colmap.py         # transform transforms.json to colmap format

create a images folder and link the train and test images into it.

mkdir -p carla_pic_0603_Town01_cam3/images
cd carla_pic_0603_Town01_cam3/images 
ln -s ../../data/carla_pic_0603_Town01/train_pic/train_camera0_00* ./ && ln -s ../../data/carla_pic_0603_Town01/train_pic/train_camera1_00* ./
ln -s ../../data/carla_pic_0603_Town01/train_pic/train_camera2_00* ./ && ln -s ../../data/carla_pic_0603_Town01/test_pic/offset_left_0m/eval_camera0_00* ./
cd ../../

GaussianPro & 3D-GS training

cd carla_pic_0603_Town01_cam1
CUDA_VISIBLE_DEVICES=0 python train.py -s carla_pic_0603_Town01_cam1 -m ./carla_pic_0603_Town01_cam1/output --position_lr_init 0.000016 --scaling_lr 0.001 --percent_dense 0.0005 --port 1021 --eval

Acknowledgement

The authors would like to thank Yuanyuan Gao and Jingfeng Li for preparing the experiments.

Citation

@article{li2024xld,
    title={XLD: A Cross-Lane Dataset for Benchmarking Novel Driving View Synthesis},
    author={Hao Li and Ming Yuan and Yan Zhang and Chenming Wu and Chen Zhao and Chunyu Song and Haocheng Feng and Errui Ding and Dingwen Zhang and Jingdong Wang},
    journal={arXiv},
    year={2024}
}