ispc-lab / LiDAR4D

💫 [CVPR 2024] LiDAR4D: Dynamic Neural Fields for Novel Space-time View LiDAR Synthesis

LiDAR4D: Dynamic Neural Fields for Novel Space-time View LiDAR Synthesis

Zehan Zheng, Fan Lu, Weiyi Xue, Guang Chen†, Changjun Jiang († Corresponding author)
CVPR 2024

Paper (arXiv) | Paper (CVPR) | Project Page | Video | Poster | Slides

This repository is the official PyTorch implementation for LiDAR4D.

Table of Contents

1. Changelog
2. Demo
3. Introduction
4. Getting started
5. Results
6. Simulation
7. Citation

Changelog

2024-6-1:🕹️ We release the simulator for easier rendering and manipulation. Happy Children's Day and Have Fun!
2024-5-4:📈 We update flow fields and improve temporal interpolation.
2024-4-13:📈 We update U-Net of LiDAR4D for better ray-drop refinement.
2024-4-5:🚀 Code of LiDAR4D is released.
2024-4-4:🔥 You can reach the preprint paper on arXiv as well as the project page.
2024-2-27:🎉 Our paper is accepted by CVPR 2024.

Demo

LiDAR4D_demo.mp4

Introduction

LiDAR4D is a differentiable LiDAR-only framework for novel space-time LiDAR view synthesis, which reconstructs dynamic driving scenarios and generates realistic LiDAR point clouds end-to-end. It adopts 4D hybrid neural representations and motion priors derived from point clouds for geometry-aware and time-consistent large-scale scene reconstruction.

Getting started

🛠️ Installation

git clone https://github.com/ispc-lab/LiDAR4D.git
cd LiDAR4D

conda create -n lidar4d python=3.9
conda activate lidar4d

# PyTorch
# CUDA 12.1
pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu121
# CUDA 11.8
# pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu118
# CUDA <= 11.7
# pip install torch==2.0.0 torchvision torchaudio

# Dependencies
pip install -r requirements.txt

# Local compile for tiny-cuda-nn
git clone --recursive https://github.com/nvlabs/tiny-cuda-nn
cd tiny-cuda-nn/bindings/torch
python setup.py install

# compile packages in utils
cd utils/chamfer3D
python setup.py install

📁 Dataset

KITTI-360 dataset (Download)

We use sequence00 (2013_05_28_drive_0000_sync) for experiments in our paper.

Download KITTI-360 dataset (2D images are not needed) and put them into data/kitti360.
(or use symlinks: ln -s DATA_ROOT/KITTI-360 ./data/kitti360/).
The folder tree is as follows:

data
└── kitti360
    └── KITTI-360
        ├── calibration
        ├── data_3d_raw
        └── data_poses

Next, run KITTI-360 dataset preprocessing: (set DATASET and SEQ_ID)

bash preprocess_data.sh

After preprocessing, your folder structure should look like this:

configs
├── kitti360_{sequence_id}.txt
data
└── kitti360
    ├── KITTI-360
    │   ├── calibration
    │   ├── data_3d_raw
    │   └── data_poses
    ├── train
    ├── transforms_{sequence_id}test.json
    ├── transforms_{sequence_id}train.json
    └── transforms_{sequence_id}val.json

🚀 Run LiDAR4D

Set corresponding sequence config path in --config and you can modify logging file path in --workspace. Remember to set available GPU ID in CUDA_VISIBLE_DEVICES.
Run the following command:

# KITTI-360
bash run_kitti_lidar4d.sh

📊 Results

KITTI-360 Dynamic Dataset (Sequences: 2350 4950 8120 10200 10750 11400)

Method	Point Cloud		Depth					Intensity
	CD↓	F-Score↑	RMSE↓	MedAE↓	LPIPS↓	SSIM↑	PSNR↑	RMSE↓	MedAE↓	LPIPS↓	SSIM↑	PSNR↑
LiDAR-NeRF	0.1438	0.9091	4.1753	0.0566	0.2797	0.6568	25.9878	0.1404	0.0443	0.3135	0.3831	17.1549
LiDAR4D (Ours) †	0.1002	0.9320	3.0589	0.0280	0.0689	0.8770	28.7477	0.0995	0.0262	0.1498	0.6561	20.0884

KITTI-360 Static Dataset (Sequences: 1538 1728 1908 3353)

Method	Point Cloud		Depth					Intensity
	CD↓	F-Score↑	RMSE↓	MedAE↓	LPIPS↓	SSIM↑	PSNR↑	RMSE↓	MedAE↓	LPIPS↓	SSIM↑	PSNR↑
LiDAR-NeRF	0.0923	0.9226	3.6801	0.0667	0.3523	0.6043	26.7663	0.1557	0.0549	0.4212	0.2768	16.1683
LiDAR4D (Ours) †	0.0834	0.9312	2.7413	0.0367	0.0995	0.8484	29.3359	0.1116	0.0335	0.1799	0.6120	19.0619

†: The latest results better than the paper.
Experiments are conducted on the NVIDIA 4090 GPU. Results may be subject to some variation and randomness.

🕹️ Simulation

After reconstruction, you can use the simulator to render and manipulate LiDAR point clouds in the whole scenario. It supports dynamic scene re-play, novel LiDAR configurations (--fov_lidar, --H_lidar, --W_lidar) and novel trajectory (--shift_x, --shift_y, --shift_z).
We also provide a simple demo setting to transform LiDAR configurations from KITTI-360 to NuScenes, using --kitti2nus in the bash script.
Check the sequence config and corresponding workspace and model path (--ckpt).
Run the following command:

bash run_kitti_lidar4d_sim.sh

The results will be saved in the workspace folder.

Acknowledgement

We sincerely appreciate the great contribution of the following works:

• tiny-cuda-nn
• LiDAR-NeRF
• NFL
• K-Planes

Citation

If you find our repo or paper helpful, feel free to support us with a star 🌟 or use the following citation:

@inproceedings{zheng2024lidar4d,
  title     = {LiDAR4D: Dynamic Neural Fields for Novel Space-time View LiDAR Synthesis},
  author    = {Zheng, Zehan and Lu, Fan and Xue, Weiyi and Chen, Guang and Jiang, Changjun},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2024}
  }

License

All code within this repository is under Apache License 2.0.