search

yunjinli / SADG-SegmentAnyDynamicGaussian

SADG: Segment Any Dynamic Gaussian Without Object Trackers
https://yunjinli.github.io/project-sadg/
16 stars 1 forks source link

readme

SADG: Segment Any Dynamic Gaussians Without Object Trackers

Project page | Paper

News

Introduction

We introduce SADG, Segment Any Dynamic Gaussian Without Object Trackers, a novel approach that combines dynamic Gaussian Splatting representation and semantic information without reliance on object IDs. We propose to learn semantically-aware features by leveraging masks generated from the Segment Anything Model (SAM) and utilizing our novel contrastive learning objective based on hard pixel mining. The learned Gaussian features can be effectively clustered without further post-processing. This enables fast computation for further object-level editing, such as object removal, composition, and style transfer by manipulating the Gaussians in the scene. Due to the lack of consistent evaluation protocol, we extend several dynamic novel-view datasets with segmentation benchmarks that allow testing of learned feature fields from unseen viewpoints. We evaluate SADG on proposed benchmarks and demonstrate the superior performance of our approach in segmenting objects within dynamic scenes along with its effectiveness for further downstream editing tasks.

teaser

Installation

## Setup the environment
git clone https://github.com/yunjinli/SADG-SegmentAnyDynamicGaussian.git
cd SADG
conda create -n SADG python=3.8 -y
pip install "git+https://github.com/facebookresearch/pytorch3d.git@stable"
pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu117
pip install opencv-python plyfile tqdm scipy wandb opencv-python scikit-learn lpips imageio[ffmpeg] dearpygui kmeans_pytorch hdbscan scikit-image bitarray
python -m pip install submodules/diff-gaussian-rasterization
python -m pip install submodules/simple-knn

## Install SAM weights
cd dependency
bash install.bash

## For enabling Text Prompt
git clone https://github.com/hkchengrex/Grounded-Segment-Anything.git
cd Grounded-Segment-Anything
export AM_I_DOCKER=False
export BUILD_WITH_CUDA=True
python -m pip install -e segment_anything
python -m pip install -e GroundingDINO

Note: If you have error from Grounding-DINO: TypeError: annotate() got an unexpected keyword argument 'labels', install Supervision to the 0.21.0 version

pip install supervision==0.21.0

Dataset Prepareation

See here

Train

See here

GUI

See here

Render

See here

Scene Editing Application

See here

Evaluation on our Mask-Benchmarks

See here

BibTex

@misc{li2024sadgsegmentdynamicgaussian,
      title={SADG: Segment Any Dynamic Gaussian Without Object Trackers},
      author={Yun-Jin Li and Mariia Gladkova and Yan Xia and Daniel Cremers},
      year={2024},
      eprint={2411.19290},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2411.19290},
}

Acknowledgement

We appreciate all the authors from 3D Gaussian Splatting,Deformable 3D Gaussians, Gaussian Grouping, SAGA for sharing their amazing works to promote further research in this area. Consider also citing their paper.

@Article{kerbl3Dgaussians,
      author       = {Kerbl, Bernhard and Kopanas, Georgios and Leimk{\"u}hler, Thomas and Drettakis, George},
      title        = {3D Gaussian Splatting for Real-Time Radiance Field Rendering},
      journal      = {ACM Transactions on Graphics},
      number       = {4},
      volume       = {42},
      month        = {July},
      year         = {2023},
      url          = {https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/}
}
@article{yang2023deformable3dgs,
    title={Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction},
    author={Yang, Ziyi and Gao, Xinyu and Zhou, Wen and Jiao, Shaohui and Zhang, Yuqing and Jin, Xiaogang},
    journal={arXiv preprint arXiv:2309.13101},
    year={2023}
}
@inproceedings{gaussian_grouping,
    title={Gaussian Grouping: Segment and Edit Anything in 3D Scenes},
    author={Ye, Mingqiao and Danelljan, Martin and Yu, Fisher and Ke, Lei},
    booktitle={ECCV},
    year={2024}
}
@article{cen2023saga,
      title={Segment Any 3D Gaussians},
      author={Jiazhong Cen and Jiemin Fang and Chen Yang and Lingxi Xie and Xiaopeng Zhang and Wei Shen and Qi Tian},
      year={2023},
      journal={arXiv preprint arXiv:2312.00860},
}