mars-nerfstudio 0.3.4

The official code repository for the paper: MARS: An Instance-aware, Modular and Realistic Simulator for Autonomous Driving

MARS: An Instance-aware, Modular and Realistic Simulator for Autonomous Driving

CICAI 2023 Best Paper Runner-up Award

For business inquiries, please contact us at [e-mail].

We have just finished a refactorization of our codebase. Now you can use pip install to start using mars instantly! Please contact us without hesitation if you encounter any issues using the latest version. Thanks!

1. Installation: Setup the environment

Prerequisites

You must have an NVIDIA video card with CUDA installed on the system. This library has been tested with version 11.7 of CUDA. You can find more information about installing CUDA here.

Create environment

Nerfstudio requires python >= 3.7. We recommend using conda to manage dependencies. Make sure to install Conda before proceeding.

conda create --name mars -y python=3.9
conda activate mars

Installation

This section will walk you through the installation process. Our system is dependent on the tiny-cuda-nn project.

pip install mars-nerfstudio
cd /path/to/tiny-cuda-nn/bindings/torch
python setup.py install

2. Training from Scratch

The following will train a MARS model.

Our repository provides dataparser for KITTI and vKITTI2 datasets, for your own data, you can write your own dataparser or convert your own dataset to the format of the provided datasets.

From Datasets

Data Preparation

The data used in our experiments should contain both the pose parameters of cameras and object tracklets. The camera parameters include the intrinsics and the extrinsics. The object tracklets include the bounding box poses, types, ids, etc. For more information, you can refer to KITTI-MOT or vKITTI2 datasets below.

KITTI

The KITTI-MOT dataset should look like this:

.(KITTI_MOT_ROOT)
├── panoptic_maps                    # (Optional) panoptic segmentation from KITTI-STEP dataset.
│   ├── colors
│   │   └── sequence_id.txt
│   ├── train
│   │   └── sequence_id
│   │       └── frame_id.png
└── training
    ├── calib
    │   └── sequence_id.txt
    ├── completion_02                # (Optional) depth completion
    │   └── sequence_id
    │       └── frame_id.png
    ├── completion_03
    │   └── sequence_id
    │       └── frame_id.png
    ├── image_02
    │   └── sequence_id
    │       └── frame_id.png
    ├── image_03
    │   └── sequence_id
    │       └── frame_id.png
    ├── label_02
    │   └── sequence_id.txt
    └── oxts
        └── sequence_id.txt

We use a monocular depth estimation model to generate the depth maps for KITTI-MOT dataset. Here is the estimation result of 0006 sequence of KITTI-MOT datasets. You can download and put them in the KITTI-MOT/training directory.

We download the KITTI-STEP annotations and generate the panoptic segmentation maps for KITTI-MOT dataset. You can download the demo panoptic maps here and put them in the KITTI-MOT directory, or you can visit the official website of KITTI-STEP for more information.

To train a reconstruction model, you can use the following command:

ns-train mars-kitti-car-depth-recon --data /data/kitti-MOT/training/image_02/0006

or if you want to use the Python script (please refer to the launch.json file in the .vscode directory):

python nerfstudio/nerfstudio/scripts/train.py mars-kitti-car-depth-recon --data /data/kitti-MOT/training/image_02/0006

vKITTI2

The vKITTI2 dataset should look like this:

.(vKITTI2_ROOT)
└── sequence_id
    └── scene_name
        ├── bbox.txt
        ├── colors.txt
        ├── extrinsic.txt
        ├── info.txt
        ├── instrinsic.txt
        ├── pose.txt
        └── frames
            ├── depth
            │   ├── Camera_0
            │   │   └── frame_id.png
            │   └── Camera_1
            │   │   └── frame_id.png
            ├── instanceSegmentation
            │   ├── Camera_0
            │   │   └── frame_id.png
            │   └── Camera_1
            │   │   └── frame_id.png
            ├── classSegmentation
            │   ├── Camera_0
            │   │   └── frame_id.png
            │   └── Camera_1
            │   │   └── frame_id.png
            └── rgb
                ├── Camera_0
                │   └── frame_id.png
                └── Camera_1
                    └── frame_id.png

To train a reconstruction model, you can use the following command:

ns-train mars-vkitti-car-depth-recon --data /data/vkitti/Scene06/clone

or if you want to use the python script:

python nerfstudio/nerfstudio/scripts/train.py mars-vkitti-car-depth-recon --data /data/vkitti/Scene06/clone

Your Own Data

For your own data, you can refer to the above data structure and write your own dataparser, or you can convert your own dataset to the format of the dataset above.

From Pre-Trained Model

Our model uses nerfstudio as the training framework, we provide the reconstruction and novel view synthesis tasks checkpoints.

Our pre-trained model is uploaded to Google Drive, you can refer to the below table to download the model.

Dataset	Scene	Setting	Start-End	Steps	PSNR	SSIM	Download	Wandb
KITTI-MOT	0006	Reconstruction	65-120	400k	27.96	0.900	model	report
	0006	Novel View Synthesis 75%	65-120	200k	27.32	0.890	model	report
	0006	Novel View Synthesis 50%	65-120	200k	26.80	0.883	model	report
	0006	Novel View Synthesis 25%	65-120	200k	25.87	0.866	model	report
Vitural KITTI-2	Scene06	Novel View Synthesis 75%	0-237	600k	32.32	0.940	model	report
	Scene06	Novel View Synthesis 50%	0-237	600k	32.16	0.938	model	report
	Scene06	Novel View Synthesis 25%	0-237	600k	30.87	0.935	model	report

You can use the following command to train a model from a pre-trained model:

ns-train mars-kitti-car-depth-recon --data /data/kitti-MOT/training/image_02/0006 --load-dir outputs/experiment_name/method_name/timestamp/nerfstudio

Model Configs

Our modular framework supports combining different architectures for each node by modifying model configurations. Here's an example of using Nerfacto for background and our category-level object model:

model=SceneGraphModelConfig(
    background_model=NerfactoModelConfig(),
    object_model_template=CarNeRFModelConfig(_target=CarNeRF),
    object_representation="class-wise",
    object_ray_sample_strategy="remove-bg",
)

If you choose to use the category-level object model, please make sure that the use_car_latents=True and the latent codes exists. We provide latent codes of some sequences on KITTI-MOT and vKITTI2 datasets here.

For more information, please refer to our provided configurations at mars/cicai_configs.py. We use wandb for logging by default, you can also specify other viewers (tensorboard/nerfstudio-viewer supported) with the --vis config. Please refer to the nerfstudio documentation for details.

Render

If you want to render with our pre-trained model, you should visit here to download our checkpoints and config. To run the render script, you need to ensure that your config is the same as the config.yml that you load in.

You can use the following command to render. You can modify output format and directory by specificing --output-format and --output-path :

python scripts/cicai_render.py --load-config outputs/kitti-recon-65-120/mars-kitti-car-depth-recon/2023-10-30_212654/config.yml --output-format video

or

python scripts/cicai_render.py --load-config outputs/kitti-recon-65-120/mars-kitti-car-depth-recon/2023-10-30_212654/config.yml --output-format images --output-path /path/to/your/output/directory

Citation

You can find our paper here. If you use this library or find the repo useful for your research, please consider citing:

@article{wu2023mars,
  author    = {Wu, Zirui and Liu, Tianyu and Luo, Liyi and Zhong, Zhide and Chen, Jianteng and Xiao, Hongmin and Hou, Chao and Lou, Haozhe and Chen, Yuantao and Yang, Runyi and Huang, Yuxin and Ye, Xiaoyu and Yan, Zike and Shi, Yongliang and Liao, Yiyi and Zhao, Hao},
  title     = {MARS: An Instance-aware, Modular and Realistic Simulator for Autonomous Driving},
  journal   = {CICAI},
  year      = {2023},
}

Acknoledgement

Part of our code is borrowed from Nerfstudio. This project is sponsored by Tsinghua-Toyota Joint Research Fund (20223930097) and Baidu Inc. through Apollo-AIR Joint Research Center.

Notice

This open-sourced version will be actively maintained and regularly updated. For more features, please contact us for a commercial version.