Arti4D: Statistical Analysis and Modeling of the Spatio-temporal Variability in Articulated 4D Shapes

ANONYMOUS AUTHOR(S)

This repository provides the implementation and demo scripts for the above paper. The codebase implements a MATLAB pipeline with Python visualization for SE(3) temporal alignment, geodesic path computation, PCA-based statistical modeling, and motion synthesis.

MATLAB:

• MATLAB R2020b or later
• Statistics and Machine Learning Toolbox
• Parallel Computing Toolbox (optional)

Python:

• Python 3.11 (recommended) or 3.10/3.9
• Python virtual environment (venv) will be created in the project directory

git clone https://github.com/Arti4D/Arti4D_code.git
cd Arti4D_code

# Create Python virtual environment (use Python 3.11)
python -m venv venv

# Activate virtual environment
# Windows:
venv\Scripts\activate
# Linux/Mac:
source venv/bin/activate

# Install dependencies
pip install -r requirements.txt

• SMPL-X Models (required for human motion data): Register and download from SMPL-X website and place in models/smplx/
  • Required files: SMPLX_MALE.npz, SMPLX_FEMALE.npz, SMPLX_NEUTRAL.npz

The repository includes demo datasets for quick testing:

• DFaust_test: Included demo dataset for human motion (SMPL-X format)

  • Located in data/DFaust_test/
  • Note: Requires SMPL-X models (see Step 3) to be downloaded and placed in models/smplx/

• Full DFaust Dataset: For full DFaust dataset, register and download from DFaust website

  • Place downloaded files in data/DFaust/
  • Data format: .npz files (SMPL-X format)
  • Note: Requires SMPL-X models (see Step 3). The batch alignment and PCA scripts work best with action-based organization. If your data is organized by subject, use data/preprocess_DFaust.m to reorganize it.

• Animal_motion Dataset: Our generated animal motion dataset (SMAL format)

  • Located in data/Animal_motion/
  • Data format: .npy files
  • Structure: Action/params_X.npy (e.g., Walking/params_1.npy, Running/params_2.npy)

The pipeline provides several scripts for different analysis tasks. Here are the recommended usage patterns:

To align one motion to another (temporal registration):

run_alignment('data/DFaust_test/50002_jumping_jacks_stageii.npz', ...
              'data/DFaust_test/50004_jumping_jacks_stageii.npz');

To see temporal warping and alignment recovery in action:

run_warping('data/DFaust_test/50002_jumping_jacks_stageii.npz');

For complete analysis pipeline (co-registration → PCA → synthesis), use:

% Animal motion (can run immediately - SMAL model included)
results = run_complete_analysis('data/Animal_motion/Walking');

% Single action (DFaust_test demo) - requires SMPL-X models
results = run_complete_analysis('data/DFaust_test');

% Multiple animal actions
results = run_complete_analysis({'data/Animal_motion/Walking', 'data/Animal_motion/Running'});

% Multiple actions (co-registration) - requires SMPL-X models
results = run_complete_analysis({'data/DFaust/chicken_wings', 'data/DFaust/jumping_jacks'});

% Custom parameters (optional - all have sensible defaults)
results = run_complete_analysis('data/DFaust_test', ...
    struct('n_pca_modes', 15, 'n_synthesis', 20));

Results: All outputs saved in timestamped directories under experiments/

• results.alignment_dir: Co-registered motions and mean motion
• results.pca_dir: PCA statistics and mode visualizations
• results.synthesis_dir: Synthesized motions

Additional scripts are available for individual steps: run_batch_alignment, run_pca, run_synthesis, run_alignment, run_warping, run_geodesic. For detailed usage and custom parameters, see the comments in each script file.