MVSGaussian

TL;DR: MVSGaussian is a Gaussian-based method designed for efficient reconstruction of unseen scenes from sparse views in a single forward pass. It also offers high-quality initialization for fast training and real-time rendering.

https://github.com/user-attachments/assets/004dd355-a9c8-477f-9c3a-1d2d11bc43dc

https://github.com/user-attachments/assets/61683f96-fa91-4d6c-81aa-d0cd84e36050

⚡ Updates

• [2024.07.16] The latest updated code supports multi-batch training (details) and inference, and a single RTX 3090 GPU is sufficient to reproduce all of our experimental results.
• [2024.07.16] Added a Demo (Custom Data) that only requires multi-view images as input.
• [2024.07.10] Code and checkpoints are released.
• [2024.07.01] Our work is accepted by ECCV2024.
• [2024.05.21] Project Page | arXiv | YouTube released.

🌟 Abstract

We present MVSGaussian, a new generalizable 3D Gaussian representation approach derived from Multi-View Stereo (MVS) that can efficiently reconstruct unseen scenes. Specifically, 1) we leverage MVS to encode geometry-aware Gaussian representations and decode them into Gaussian parameters. 2) To further enhance performance, we propose a hybrid Gaussian rendering that integrates an efficient volume rendering design for novel view synthesis. 3) To support fast fine-tuning for specific scenes, we introduce a multi-view geometric consistent aggregation strategy to effectively aggregate the point clouds generated by the generalizable model, serving as the initialization for per-scene optimization. Compared with previous generalizable NeRF-based methods, which typically require minutes of fine-tuning and seconds of rendering per image, MVSGaussian achieves real-time rendering with better synthesis quality for each scene. Compared with the vanilla 3D-GS, MVSGaussian achieves better view synthesis with less training computational cost. Extensive experiments on DTU, Real Forward-facing, NeRF Synthetic, and Tanks and Temples datasets validate that MVSGaussian attains state-of-the-art performance with convincing generalizability, real-time rendering speed, and fast per-scene optimization.

🔨 Installation

Clone our repository

git clone https://github.com/TQTQliu/MVSGaussian.git
cd MVSGaussian

Set up the python environment

conda create -n mvsgs python=3.7.13
conda activate mvsgs
pip install -r requirements.txt
pip install torch==1.13.1+cu116 torchvision==0.14.1+cu116 torchaudio==0.13.1 -f https://download.pytorch.org/whl/torch_stable.html

Install Gaussian Splatting renderer

pip install lib/submodules/diff-gaussian-rasterization
pip install lib/submodules/simple-knn

🤗 Demo (Custom Data)

Inference

First, prepare the multi-view image data, and then run colmap. Here, we take examples/scene1 (examples data) as an example:

python lib/colmap/imgs2poses.py -s examples/scene1

Tip: If you already have sparse reconstruction results, i.e. sparse/0/cameras.bin, sparse/0/images.bin, sparse/0/points3D.bin, and want to skip the colmap reconstruction step of the script, you can place the above sparse folder in the examples/scene1 directory and run the same command. The script recognizes that sparse reconstruction results already exist, automatically skips the colmap reconstruction phase, and simply organizes the existing results to produce the required poses_bounds.npy.

And execute the following command to obtain novel views:

python run.py --type evaluate --cfg_file configs/mvsgs/colmap_eval.yaml test_dataset.data_root examples test_dataset.scene scene1

or videos:

python run.py --type evaluate --cfg_file configs/mvsgs/colmap_eval.yaml test_dataset.data_root examples test_dataset.scene scene1 save_video True

For your own dataset, like:

data
  ├── scene1                   
      ├── images                
      ├── sparse   
               ├── 0
      ├── poses_bounds.npy
 ├── scene2
  ...

First modify the scenes list here and then run

python run.py --type evaluate --cfg_file configs/mvsgs/colmap_eval.yaml test_dataset.data_root examples

Train on your own data

If you want to train our model on your own dataset, first organize the dataset as follows：

data
  ├── scene1                   
      ├── images                
      ├── sparse   
               ├── 0
      ├── poses_bounds.npy
 ├── scene2
  ...

Modify the scenes list here and then execute the following commands:

python train_net.py --cfg_file configs/mvsgs/colmap_eval.yaml train_dataset.data_root examples test_dataset.data_root examples

You can specify the gpus in configs/mvsgs/dtu_pretrain.yaml. And you can modify the exp_name in the configs/mvsgs/dtu_pretrain.yaml. Before training, the code will first check whether there is checkpoint in trained_model/mvsgs/exp_name, and if so, the latest checkpoint will be loaded. During training, the tensorboard log will be save in record/mvsgs/exp_name, the trained checkpoint will be save in trained_model/mvsgs/exp_name, and the rendering results will be saved in result/mvsgs/exp_name.

Per-scene optimization

For per-scene optimization, first run the generalizable model to obtain the point cloud as initialization for subsequent optimization.

python run.py --type evaluate --cfg_file configs/mvsgs/colmap_eval.yaml test_dataset.data_root examples test_dataset.scene scene1 save_ply True dir_ply <path to save ply>

The point cloud will be saved in <path to save ply>/scene1/scene1.ply. Note that this point cloud is a normal geometric point cloud, not a Gaussian point cloud, and you can open it through MeshLab.

And then run the 3DGS optimization:

python lib/train.py  --eval --iterations <iter> -s examples/scene1 -p <path to save ply>

The optimized Gaussian point cloud will be saved in output/scene1/point_cloud/iteration_<iter>/point_cloud.ply, and you can open it through 3DGS viewer.

Run the following commands to synthesize target views and calculate metrics:

python lib/render.py -c -m output/scene1 --iteration <iter> -p <path to save ply>
python lib/metrics.py -m output/scene1

Add -v to obtain the rendered video:

python lib/render.py -c -m output/scene1 -p <path to save ply> -v

📦 Datasets

• DTU

  Download DTU data and Depth raw. Unzip and organize them as:

  mvs_training
      ├── dtu                   
          ├── Cameras                
          ├── Depths   
          ├── Depths_raw
          └── Rectified
  

• Download NeRF Synthetic, Real Forward-facing, and Tanks and Temples datasets.

🚂 Training

Train generalizable model

To train a generalizable model from scratch on DTU, specify data_root in configs/mvsgs/dtu_pretrain.yaml first and then run:

python train_net.py --cfg_file configs/mvsgs/dtu_pretrain.yaml train.batch_size 4

You can specify the gpus in configs/mvsgs/dtu_pretrain.yaml. And you can modify the exp_name in the configs/mvsgs/dtu_pretrain.yaml. Before training, the code will first check whether there is checkpoint in trained_model/mvsgs/exp_name, and if so, the latest checkpoint will be loaded. During training, the tensorboard log will be save in record/mvsgs/exp_name, the trained checkpoint will be save in trained_model/mvsgs/exp_name, and the rendering results will be saved in result/mvsgs/exp_name.

Our code also supports multi-gpu training. The released pretrained model (paper) was trained with 4 RTX 3090 GPUs with a batch size of 1 for each GPU:

python -m torch.distributed.launch --nproc_per_node=4 train_net.py --cfg_file configs/mvsgs/dtu_pretrain.yaml distributed True gpus 0,1,2,3 train.batch_size 1

You can also use 4 GPUs, with a batch size of 4 for each GPU:

python -m torch.distributed.launch --nproc_per_node=4 train_net.py --cfg_file configs/mvsgs/dtu_pretrain.yaml distributed True gpus 0,1,2,3 train.batch_size 4

We provide the results as a reference below:

Per-scene optimization

One strategy is to optimize only the initial Gaussian point cloud provided by the generalizable model.

bash scripts/mvsgs/llff_ft.sh
bash scripts/mvsgs/nerf_ft.sh
bash scripts/mvsgs/tnt_ft.sh

We provide optimized Gaussian point clouds for each scenes here.

You can also run the following command to get the results of vanilla 3D-GS, whose initialization is obtained via COLMAP.

bash scripts/3dgs/llff_ft.sh
bash scripts/3dgs/nerf_ft.sh
bash scripts/3dgs/tnt_ft.sh

It is worth noting that for the LLFF dataset, the point cloud in the original dataset is obtained by using all views. For fair comparison, we only use the training view set to regain the point cloud, so we recommend downloading the LLFF dataset we processed.

(Optional) Another approach is to optimize the entire pipeline, similar to NeRF-based methods.

Here we take the fern on the LLFF as an example:

cd ./trained_model/mvsgs
mkdir llff_ft_fern
cp dtu_pretrain/latest.pth llff_ft_fern
cd ../..
python train_net.py --cfg_file configs/mvsgs/llff/fern.yaml

🎯 Evaluation

Evaluation on DTU

Download the pretrained model and put it into trained_model/mvsgs/dtu_pretrain/latest.pth

Use the following command to evaluate the pretrained model on DTU:

python run.py --type evaluate --cfg_file configs/mvsgs/dtu_pretrain.yaml mvsgs.cas_config.render_if False,True mvsgs.cas_config.volume_planes 48,8 mvsgs.eval_depth True

The rendered images will be saved in result/mvsgs/dtu_pretrain.

Evaluation on Real Forward-facing

python run.py --type evaluate --cfg_file configs/mvsgs/llff_eval.yaml

Evaluation on NeRF Synthetic

python run.py --type evaluate --cfg_file configs/mvsgs/nerf_eval.yaml

Evaluation on Tanks and Temples

python run.py --type evaluate --cfg_file configs/mvsgs/tnt_eval.yaml

Render videos

Add the save_video True argument to save videos, such as:

python run.py --type evaluate --cfg_file configs/mvsgs/llff_eval.yaml save_video True

For optimized Gaussians, add -v to save videos, such as:

python lib/render.py -m output/$scene -p $dir_ply -v

See scripts/mvsgs/nerf_ft.sh for $scene and $dir_ply.

📝 Citation

If you find our work useful for your research, please cite our paper.

@inproceedings{liu2024mvsgaussian,
  title={Mvsgaussian: Fast generalizable gaussian splatting reconstruction from multi-view stereo},
  author={Liu, Tianqi and Wang, Guangcong and Hu, Shoukang and Shen, Liao and Ye, Xinyi and Zang, Yuhang and Cao, Zhiguo and Li, Wei and Liu, Ziwei},
  booktitle={European Conference on Computer Vision},
  pages={37--53},
  year={2024},
  organization={Springer}
}

🔗 Related Papers

You may also be interested in our related work on generalizable NeRF and 3DGS:

• [CVPR 2024] Geometry-aware Reconstruction and Fusion-refined Rendering for Generalizable Neural Radiance Fields

• [ICCV 2025] MuGS: Multi-Baseline Generalizable Gaussian Splatting Reconstruction

😃 Acknowledgement

This project is built on source codes shared by Gaussian-Splatting, ENeRF, MVSNeRF, and LLFF. Many thanks for their excellent contributions!

📧 Contact

If you have any questions, please feel free to contact Tianqi Liu (tq_liu at hust.edu.cn).