MVSplat: Efficient 3D Gaussian Splatting
from Sparse Multi-View Images

Yuedong Chen  ·  Haofei Xu  ·  Chuanxia Zheng  ·  Bohan Zhuang
Marc Pollefeys  ·  Andreas Geiger  ·  Tat-Jen Cham  ·  Jianfei Cai

ECCV 2024 Oral

Paper | Project Page | Pretrained Models

• 20/01/25 Update: Check out Cheng's PanSplat [CVPR '25], which extends MVSplat to higher resolutions (up to 4K) and highlights the use of a hierarchical spherical cost volume and two-step deferred backpropagation for memory-efficient training.
• 08/11/24 Update: Explore our MVSplat360 [NeurIPS '24], an upgraded MVSplat that combines video diffusion to achieve 360° NVS for large-scale scenes from just 5 input views!
• 21/10/24 Update: Check out Haofei's DepthSplat [CVPR '25] if you are interested in feed-forward 3DGS on more complex scenes (DL3DV-10K) and more input views (up to 12 views)!

https://github.com/donydchen/mvsplat/assets/5866866/c5dc5de1-819e-462f-85a2-815e239d8ff2

Installation

To get started, clone this project, create a conda virtual environment using Python 3.10+, and install the requirements:

git clone https://github.com/donydchen/mvsplat.git
cd mvsplat
conda create -n mvsplat python=3.10
conda activate mvsplat
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

Acquiring Datasets

RealEstate10K and ACID

Our MVSplat uses the same training datasets as pixelSplat. Below we quote pixelSplat's detailed instructions on getting datasets.

pixelSplat was trained using versions of the RealEstate10k and ACID datasets that were split into ~100 MB chunks for use on server cluster file systems. Small subsets of the Real Estate 10k and ACID datasets in this format can be found here. To use them, simply unzip them into a newly created datasets folder in the project root directory.

If you would like to convert downloaded versions of the Real Estate 10k and ACID datasets to our format, you can use the scripts here. Our preprocessed versions of the datasets can be found here.

DTU (For Testing Only)

• Download the preprocessed DTU data dtu_training.rar.
• Convert DTU to chunks by running python src/scripts/convert_dtu.py --input_dir PATH_TO_DTU --output_dir datasets/dtu
• [Optional] Generate the evaluation index by running python src/scripts/generate_dtu_evaluation_index.py --n_contexts=N, where N is the number of context views. (For N=2 and N=3, we have already provided our tested version under /assets.)

Running the Code

Evaluation

To render novel views and compute evaluation metrics from a pretrained model,

• get the pretrained models, and save them to /checkpoints

• run the following:

# re10k
python -m src.main +experiment=re10k \
checkpointing.load=checkpoints/re10k.ckpt \
mode=test \
dataset/view_sampler=evaluation \
test.compute_scores=true

# acid
python -m src.main +experiment=acid \
checkpointing.load=checkpoints/acid.ckpt \
mode=test \
dataset/view_sampler=evaluation \
dataset.view_sampler.index_path=assets/evaluation_index_acid.json \
test.compute_scores=true

• the rendered novel views will be stored under outputs/test

To render videos from a pretrained model, run the following

# re10k
python -m src.main +experiment=re10k \
checkpointing.load=checkpoints/re10k.ckpt \
mode=test \
dataset/view_sampler=evaluation \
dataset.view_sampler.index_path=assets/evaluation_index_re10k_video.json \
test.save_video=true \
test.save_image=false \
test.compute_scores=false

Training

Run the following:

# download the backbone pretrained weight from unimatch and save to 'checkpoints/'
wget 'https://s3.eu-central-1.amazonaws.com/avg-projects/unimatch/pretrained/gmdepth-scale1-resumeflowthings-scannet-5d9d7964.pth' -P checkpoints
# train mvsplat
python -m src.main +experiment=re10k data_loader.train.batch_size=14

Our models are trained with a single A100 (80GB) GPU. They can also be trained on multiple GPUs with smaller RAM by setting a smaller data_loader.train.batch_size per GPU.

#SBATCH --nodes=2           # should match with trainer.num_nodes
#SBATCH --gres=gpu:2        # gpu per node
#SBATCH --ntasks-per-node=2

# optional, for debugging
export NCCL_DEBUG=INFO
export HYDRA_FULL_ERROR=1
# optional, set network interface, obtained from ifconfig
export NCCL_SOCKET_IFNAME=[YOUR NETWORK INTERFACE]
# optional, set IB GID index
export NCCL_IB_GID_INDEX=3

# run the command with 'srun'
srun python -m src.main +experiment=re10k \
data_loader.train.batch_size=4 \
trainer.num_nodes=2

python -m src.main +experiment=re10k data_loader.train.batch_size=14 \
checkpointing.load=checkpoints/re10k.ckpt \
checkpointing.resume=false

Ablations

We also provide a collection of our ablation models (under folder 'ablations'). To evaluate them, e.g., the 'base' model, run the following command

# Table 3: base
python -m src.main +experiment=re10k \
checkpointing.load=checkpoints/ablations/re10k_worefine.ckpt \
mode=test \
dataset/view_sampler=evaluation \
test.compute_scores=true \
wandb.name=abl/re10k_base \
model.encoder.wo_depth_refine=true 

Cross-Dataset Generalization

We use the default model trained on RealEstate10K to conduct cross-dataset evaluations. To evaluate them, e.g., on DTU, run the following command

# Table 2: RealEstate10K -> DTU
python -m src.main +experiment=dtu \
checkpointing.load=checkpoints/re10k.ckpt \
mode=test \
dataset/view_sampler=evaluation \
dataset.view_sampler.index_path=assets/evaluation_index_dtu_nctx2.json \
test.compute_scores=true

More running commands can be found at more_commands.sh.

BibTeX

@article{chen2024mvsplat,
    title   = {MVSplat: Efficient 3D Gaussian Splatting from Sparse Multi-View Images},
    author  = {Chen, Yuedong and Xu, Haofei and Zheng, Chuanxia and Zhuang, Bohan and Pollefeys, Marc and Geiger, Andreas and Cham, Tat-Jen and Cai, Jianfei},
    journal = {arXiv preprint arXiv:2403.14627},
    year    = {2024},
}

Acknowledgements

The project is largely based on pixelSplat and has incorporated numerous code snippets from UniMatch. Many thanks to these two projects for their excellent contributions!