SCube: Instant Large-Scale Scene Reconstruction using VoxSplats

SCube: Instant Large-Scale Scene Reconstruction using VoxSplats
Xuanchi Ren$^* $, Yifan Lu$^* $, Hanxue Liang, Zhangjie Wu, Huan Ling, Mike Chen, Sanja Fidler, Francis Williams, Jiahui Huang
* indicates equal contribution
[Project Page]

News

• 2024-12-11: Also check out our latest research InfiniCube.
• 2024-12-11: Code released! We have also integrated XCube into this repository for easier use. For detailed instructions, please refer to the XCube Instruction.

Installation

mkdir SCube-release # wrap it in a folder
cd SCube-release
git clone https://github.com/nv-tlabs/SCube.git 
mkdir wandb
mkdir waymo_tfrecords 
cd SCube

(Optional) Install libMamba for a huge quality of life improvement when using Conda

conda update -n base conda
conda install -n base conda-libmamba-solver
conda config --set solver libmamba

Install the conda environment

conda env create -f environment.yml
conda activate scube
mim install "mmcv>=2.0.0"
pip install "mmsegmentation>=1.0.0"

Setup WandB

Our training script highly relies on WandB. Please register an account for WandB first and get your API_key. Then you can setup for your machine by running this command in the terminal:

wandb login # requires your API key

Data Preparation

Data Split Downloading

First download our splits of waymo dataset.

gdown https://drive.google.com/drive/folders/1d8V4aoHsDf5U0-SH_VY2MOZJdYvOG8TM -O ../waymo_split --folder

Waymo Dataset Downloading

Download the all the training & validation clips from waymo perception dataset v1.4.2 to the SCube-release/waymo_tfrecords.

If you have sudo, you can use gcloud to download them from terminal.

gcloud installation (need sudo) and downloading from terminal

sudo apt-get update
sudo apt-get install apt-transport-https ca-certificates gnupg curl
curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | sudo gpg --dearmor -o /usr/share/keyrings/cloud.google.gpg
echo "deb [signed-by=/usr/share/keyrings/cloud.google.gpg] https://packages.cloud.google.com/apt cloud-sdk main" | sudo tee -a /etc/apt/sources.list.d/google-cloud-sdk.list
sudo apt-get update && sudo apt-get install google-cloud-cli

Then you can login your google account and download the above tfrecords via

# or use `gcloud init --no-launch-browser` if you are in a remote terminal session
gcloud init 
bash datagen/download_waymo.sh datagen/waymo_all.json ../waymo_tfrecords

After downloading tfrecord files, we expect a folder structure as follows:

SCube-release
|-- SCube
|   `-- ...
|-- waymo_split
|   `-- ...
`-- waymo_tfrecords
    |-- segment-10247954040621004675_2180_000_2200_000_with_camera_labels.tfrecord
    |-- segment-11379226583756500423_6230_810_6250_810_with_camera_labels.tfrecord
    |-- ...
    `-- segment-1172406780360799916_1660_000_1680_000_with_camera_labels.tfrecord

Note

If you download the tfrecord files from the console, you will have prefixes like individual_files_training_ or individual_files_validation_. Make sure these prefixes are removed before further processing.

Data Processing

We first extract the image, LiDAR, pose, intrinsic, annotation files out of tfrecords, which will be saved in ../waymo_ns. If you have multiple CPU cores, you can set --num_workers to your core number for parallelization.

python datagen/waymo2nerfstudio.py -i ../waymo_tfrecords -o ../waymo_ns [--num_workers YOUR_CPU_CORE]

We then process the raw data into webdataset format required by SCube training while infering SegFormer for sky mask and Metric3Dv2 for GT depth. The results will be saved in ../waymo_webdataset. Note that this step is very time-comsuming for 1000 clips x 198 frames x 5 cameras, which can take over 1 day on 8x A100 GPUs.

Note

You might encounter AssertionError: MMCV==2.2.0 is used but incompatible. Please install mmcv>=2.0.0rc4. when using mmcv. Solution here.

Note

You might encounter RuntimeError: Failed to find function: mono.model.backbones.vit_large_reg when using Metric3D v2. Solution here

# single GPU
python datagen/nerfstudio2webdataset.py -i ../waymo_ns -o ../waymo_webdataset

# multiple GPU
torchrun --nnodes=1 --nproc-per-node=8 'datagen/nerfstudio2webdataset.py' -i ../waymo_ns -o ../waymo_webdataset

Now the ../waymo_ns and ../waymo_tfrecords will not be used anymore. We expect the folder structure like:

SCube-release
|-- SCube
|   `-- ...
|-- waymo_split
|   `-- ...
`-- waymo_webdataset
    `-- ...

GT Voxel Downloading

We provide the webdataset-format files of ground-truth voxels in Google Drive (need request). After getting approval, download and uncompress it in SCube-release/waymo_webdataset along with other attributes.

Model Training

VAE Training

# Coarse stage VAE
python train.py configs/waymo_scube/train_vae_64x64x64_dense_height_down2_residual.yaml --wname train_vae_64x64x64_dense_height_down2_residual --max_epochs 10 --gpus 8 --eval_interval 1

# Fine stage VAE
python train.py configs/waymo_scube/train_vae_256x256x128_sparse.yaml --wname train_vae_256x256x128_sparse --max_epochs 10 --gpus 8 --eval_interval 1

Geometry Reconstruction Training (Latent Voxel Diffusion Model)

Important

You need to modify the vae_checkpoint in configs/waymo_scube/train_diffusion_64x64x64_image_cond.yaml and configs/waymo_scube/train_diffusion_256x256x128_sparse.yaml to your own experiments. In you wandb webpage, you can find a unique "run id" for your experiment and replace the [YOUR_RUN_ID] in the yaml file.

# Image conditioned voxel diffusion
python train.py configs/waymo_scube/train_diffusion_64x64x64_image_cond.yaml --wname train_diffusion_64x64x64_image_cond --max_epochs 40 --gpus 8 --eval_interval 1

# Semantic conditioned upsampling voxel diffusion
python train.py configs/waymo_scube/train_diffusion_256x256x128_sparse.yaml --wname train_diffusion_256x256x128_sparse --max_epochs 20 --gpus 8 --eval_interval 1

Appearance Reconstruction Training (GSM, Gaussian Splatting Model)

python train.py configs/waymo_scube/train_gsm_unet3d_view3.yaml --wname train_gsm_unet3d_view3 --max_epochs 30 --gpus 8 --eval_interval 1

Model Inference

Inference VAE

# replace nvidia-toronto to your own wandb account!
python inference/vae.py none --ckpt_vae wdb:nvidia-toronto/scube-scene-recon/waymo_wds/train_vae_64x64x64_dense_height_down2_residual
python inference/vae.py none --ckpt_vae wdb:nvidia-toronto/scube-scene-recon/waymo_wds/train_vae_256x256x128_sparse

This will generate a folder ../vae_output_waymo_wds/, storing predicted & GT grid + semantics. Specify --ckpt_vae to the experiment name.

Inference Diffusion

# replace nvidia-toronto to your own wandb account!
python inference/diffusion.py none --use_ema --use_ddim --ddim_step 100 --ckpt_dm wdb:nvidia-toronto/scube-scene-recon/waymo_wds/train_diffusion_64x64x64_image_cond
python inference/diffusion.py none --use_ema --use_ddim --ddim_step 100 --ckpt_dm wdb:nvidia-toronto/scube-scene-recon/waymo_wds/train_diffusion_256x256x128_sparse

This will generate a folder ../diffusion_output_waymo_wds/, storing predicted & GT voxel grid with semantics. Specify --ckpt_dm to the experiment name.

Inference GSM

# replace nvidia-toronto to your own wandb account!
python inference/gaussian_feedforward.py none --ckpt_gsm wdb:nvidia-toronto/scube-scene-recon/waymo_wds/train_gsm_unet3d_view3 [--save_gs]

This will generate a folder ../diffusion_output_waymo_wds/, storing predicted & GT grid + semantics. Specify --ckpt_gsm to the experiment name. Add [--save_gs] to save the 3D Gaussians.

Inference Full Pipeline

# replace nvidia-toronto to your own wandb account!
python inference/diffusion_cascading_gsm.py none \
    --use_ema --use_ddim --ddim_step 100 \
    --ckpt_dm_c wdb:nvidia-toronto/scube-scene-recon/waymo_wds/train_diffusion_64x64x64_image_cond \
    --ckpt_dm_f wdb:nvidia-toronto/scube-scene-recon/waymo_wds/train_diffusion_256x256x128_sparse \
    --ckpt_gsm wdb:nvidia-toronto/scube-scene-recon/waymo_wds/train_gsm_unet3d_view3 \
    --input_frame_offsets [0] \
    --sup_frame_offsets [0,5,10] \
    --split val \
    --val_starting_frame 0

Visualize the Inference Result

Visualize Grids

In ../vae_output_waymo_wds/ and ../diffusion_output_waymo_wds/ and ../cascading_diffusion_output_waymo_wds, you can find some folders containing x.pt and x_gt.pt, where x is the sample index for your inference. Run the following command to launch a interactive visualizer (even on a remote-ssh machine) in a web browser.

python inference/visualize_grid_compare.py -p <PT_FILES_FOLDER> [-t voxel]

where -t is available for voxel or pc. Do not use voxel for the fine stage, which is very heavy for rendering. It would be easy to use it on a remote machine with VSCode; VSCode will forward the port automatically for you.

---

Visualize 3DGS

In ../splat_output_waymo_wds and cascading_diffusion_output_waymo_wds, you can find some folder containing x_splat.pkl, where x is the sample index for your inferenece. You can visualize the 3DGS with

python inference/visualize_3dgs_pkl.py -p <3DGS_PKL_FILE>

---

Shortcut

You could install this project to create short cut for visualization.

poetry install 

Then you will have command vis-pair equivalent to python inference/visualize_grid_compare.py and vis-gs equivalent to python inference/visualize_3dgs_pkl.py. You can run visualization as

vis-pair -p <PT_FILES_FOLDER>
vis-gs -p <3DGS_PKL_FILE>

License

Copyright © 2024, NVIDIA Corporation & affiliates. All rights reserved. This work is made available under the Nvidia Source Code License.

Related Works

• Lu et al. 2024. InfiniCube: Unbounded and Controllable Dynamic 3D Driving Scene Generation with World-Guided Video Models.
• Ren et al. 2024. XCube: Large-Scale 3D Generative Modeling using Sparse Voxel Hierarchies.
• Williams et al. 2024. 𝑓VDB: A Deep-Learning Framework for Sparse, Large-Scale, and High-Performance Spatial Intelligence.

Acknowledgement

This repo is based on https://github.com/nv-tlabs/XCube.

Citation

@inproceedings{
    ren2024scube,
    title={SCube: Instant Large-Scale Scene Reconstruction using VoxSplats},
    author={Ren, Xuanchi and Lu, Yifan and Liang, Hanxue and Wu, Jay Zhangjie and Ling, Huan and Chen, Mike and Fidler, Sanja annd Williams, Francis and Huang, Jiahui},
    booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
    year={2024},
}