LGS: A Light-weight 4D Gaussian Splatting for Efficient Surgical Scene Reconstruction

Project Page | ArXiv Paper

Accepted by International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2024 poster)

Hengyu Liu^*, Yifan Liu^*, Chenxin Li^*, Wuyang Li, Yixuan Yuan^✉

The Chinese University of Hong Kong

^* Equal Contribution, ^✉ Corresponding Author.

💡Highlight

• We propose a holistic Lightweight 4D Gaussian Splatting (LGS) framework that allows for achieving satisfactory endoscopic reconstruction with both efficient rendering and storing.
• We present a Deformation-Aware Pruning (DAP) which alleviates the Quantity burden of Gaussian representation.
• We propose a Gaussian-Attribute Pruning (GAP), which addresses the High-dimension burden of Gaussian attributes.
• We present Feature Field Condensation (FFC) which mitigates the High-resolution burden of spatial-temporal deformable fields.
• Experimental results show that LGS can achieve higher storage efficiency with an over $9\times$ compression rate, whilst maintaining pleasing reconstruction quality and rendering speed.

🛠Setup

• submodules/simple_knn can be download from here
• compress-diff-gaussian-rasterization can be download from compress-diff-gaussian-rasterization

git clone https://github.com/CUHK-AIM-Group/LGS.git
cd LGS
conda create -n lgs python=3.7
conda activate lgs

pip install -r requirements.txt

pip install -e submodules/depth-diff-gaussian-rasterization
pip install -e submodules/simple-knn

Tips: 2 GB GPU memory is required for training and inference.

📚Data Preparation

Same to the 📚Data Preparation process of EndoGaussian:

ENDONERF The dataset provided in EndoNeRF is used. You can download and process the dataset from their website. We use the two accessible clips including 'pulling_soft_tissues' and 'cutting_tissues_twice'.

SCARED The dataset provided in SCARED is used. To obtain a link to the data and code release, sign the challenge rules and email them to [email protected]. You will receive a temporary link to download the data and code. Follow MICCAI_challenge_preprocess to extract data. The resulted file structure is as follows.

The file structure is as follows.

├── data
│   | endonerf 
│     ├── pulling
│     ├── cutting 
│   | scared
│     ├── dataset_1
│       ├── keyframe_1
│           ├── data
│       ├── ...
│     ├── dataset_2
|     ├── ...

⏳Training LGS

Run the following code to train LGS in a distillation scheme.

cd to this repo
conda activate engs

dataset="name/of/dataset"
test_name="name/of/output/file"
config_name="name/of/config"

check_point="/path/to/teacher/model/chkpntN.pth"
deformation_path="/path/to/teacher/model/point_cloud/iteration_N"

CUDA_VISIBLE_DEVICES=GPU_ID python train.py \
    -s /path/to/endonerf/cutting_tissues_twice/ \
    -m /path/to/endonerf/endonerf-output/ \
    --expname $dataset/new_init/$test_name \
    --start_checkpoint $check_point \
    --deformatioin_model_path $deformation_path \
    --teacher_model $check_point \
    --enable_covariance \
    --new_max_sh 2 \
    --iteration 4000 \
    --configs $config_name \
    --prune_percent 0.3 \
    --prune_decay 0.6 \
    --prune_iterations 2000 \
    --prune \
    --prune_threshold 0.5 \
    --prune_sh \
    --prune_deform \
    --v_pow 0.1 \
    --gt \
    --port $port > /path/to/log

Notice:

• (1) Teacher Model is a trained EndoGaussian model, while the size of the deformation filed of teacher model should be set the same as LGS:

kplanes_config = {
  'grid_dimensions': 2,
  'input_coordinate_dim': 4,
  'output_coordinate_dim': 64,
  'resolution': [64, 64, 64, 100]
}

• (2) While training EndoGaussian, EndoGaussian's two saving processs are at different times, one before pruning the Gaussians and the other after pruning. Please refer to here and get the consistent models.

📏Metric Evaluation

Run the following code to render and get the evaluaton.

CUDA_VISIBLE_DEVICES=GPU_ID python render.py --model_path /path/to/trained/model \
    --skip_train \
    --new_max_sh 2 \
    --prune_deform \
    --configs $config_name > /path/to/log

CUDA_VISIBLE_DEVICES=GPU_ID python metrics.py --model_path /path/to/trained/model > /path/to/log

🎪Experiment Results

Quantitative Results

Qualitative Results

🛒TODO List

• [X] Release code for LGS
• [X] Clean up the code for LGS
• [ ] Updae the code for ablation experiments

🎈Acknowledgements

Greatly appreciate the tremendous effort for the following projects!

• LightGaussian
• EndoGaussian
• 3D GS
• 4D GS
• EndoNeRF
• compress-diff-gaussian-rasterization

📜Citation

If you find this work helpful for your project,please consider citing the following paper:

@article{liu2024lgs,
  title={LGS: A Light-weight 4D Gaussian Splatting for Efficient Surgical Scene Reconstruction},
  author={Liu, Hengyu and Liu, Yifan and Li, Chenxin and Li, Wuyang and Yuan, Yixuan},
  journal={arXiv preprint arXiv:2406.16073},
  year={2024}
}