GEAR: GEometry-motion Alternating Refinement for Articulated Object Modeling with Gaussian Splatting

This is the official code repository for our paper GEAR: GEometry-motion Alternating Refinement for Articulated Object Modeling with Gaussian Splatting, accepted at CVPR 2026 (Findings).

Method Overview

Demo

Environment setup

We recommend Python 3.10 on Linux with an NVIDIA GPU and a CUDA stack that matches your PyTorch build.

1. PyTorch — Install torch and torchvision for your CUDA version from the official install guide before installing other dependencies.

2. Python dependencies — From the repository root:

   pip install -r requirements.txt

   The file includes Segment Anything via Git; pip needs network access to GitHub.

3. GS extensions (required) — After PyTorch is installed, build and install the local packages (still from the repo root):

   pip install ./submodules/simple-knn
   pip install ./submodules/diff-surfel-rasterization
   pip install ./utils/pointnet_lib

   If you use Git submodules, initialize simple-knn first, for example:

   git submodule update --init submodules/simple-knn

4. SAM checkpoint — Mask extraction (utils/get_mask.py, stage 1) expects the ViT-H weights at:

   submodules/sam_vit_h_4b8939.pth
   

   Download (example):

   wget -O submodules/sam_vit_h_4b8939.pth \
     https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth

   You can override the path with --sam_checkpoint when running utils/get_mask.py.

5. PyTorch3D and tiny-cuda-nn — If installation fails, see PyTorch3D install and the tiny-cuda-nn build notes (CUDA toolkit / nvcc required for extensions).

Dataset

We introduce the GEAR_Multi dataset for articulated object modeling. You can download the dataset from Google Drive here: 👉 Download GEAR_Multi Dataset

Code Structure & Usage

The pipeline is primarily divided into five stages. We provide bash scripts in the scripts/ directory to run each stage sequentially. Make sure to adjust the dataset paths and variables inside the scripts according to your local setup before running them.

1. Extract Masks

Extract 2D masks for the articulated object parts using SAM (Segment Anything Model) based on the initial images.

bash scripts/1_get_mask.sh

2. Coarse Optimization

Train a coarse Gaussian Splatting model to obtain a global geometry initialization.

bash scripts/2_coarse.sh

3. Voxelization and Dynamic Joint Extraction

Utilize the geometry from the coarse model to perform Top-K connected component voxelization. This step extracts dynamic joints and initializes their coarse poses.

bash scripts/3_voxelize.sh

4. Fine Training (GEAR)

Perform the Geometry-motion Alternating Refinement. This phase finely optimizes the joint parameters (axis, rotation, translation) alongside the 3D Gaussian attributes.

bash scripts/4_train.sh

5. Rendering & Evaluation

Render the novel views, visualize the part segmentation, and evaluate the resulting geometric and visual metrics.

bash scripts/5_render.sh

Citation

If you find our work or dataset useful, please consider citing:

@misc{gear2026,
      title={GEAR: GEometry-motion Alternating Refinement for Articulated Object Modeling with Gaussian Splatting}, 
      author={Jialin Li and Bin Fu and Ruiping Wang and Xilin Chen},
      year={2026},
      eprint={2604.07728},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.07728}, 
}