SMPL on terrain (SCENIC Dataset)

Pairing up your human motion with terrains to build height-adaptive motion models? This repository contains the SCENIC dataset for motion-scene-text data sequences, enabling research in motion modeling and scene-aware control.

Dataset Overview

You can download the necessary components from the following links:

• SCENIC Dataset: Download here
• Terrain Asset: Download here

Breakdown of the Dataset:

• npz/: Contains original motion sequences in the SMPL format.
• pkl/: Contains the split motion segments used for fitting to terrains.
• text/: Contains text annotations for various styles of motion for each motion segment.

Using the Dataset

1. To preprocess and visualize the dataset, run:

   python process_and_visualize.py

2. The paired terrains will be saved as watertight meshes in the scenes/ directory.

Credit

Credit to Phase-functioned Neural Network (PFNN) for providing the terrain asset and the fitting mechanism.

Citation

If you use the SCENIC dataset or our code in your research, please cite us:

@article{zhang2024scenic,
    title = {SCENIC: Scene-aware Semantic Navigation with Instruction-guided Control},
    author = {Zhang, Xiaohan and Starke, Sebastian and Guzov, Vladimir and Dhamo, Helisa and Pérez Pellitero, Eduardo and Pons-Moll, Gerard},
    booktitle = {Arxiv},
    year = {2024},
}

@article{pfnn,
    author = {Daniel Holden and Taku Komura and Jun Saito},
    title = {Phase-functioned neural networks for character control},
    journal = {{ACM} Trans. Graph.},
    year = {2017}
}