- Code and model for training and testing.
- Full-size synthetic dataset for retraining.
- Full-size real-world dataset for testing.
- Blender files for generating synthetic data.
Git clone this repo:
git clone -b main --single-branch https://github.com/WenbWa/TRTM.git
cd TRTM
Init conda environment:
conda create -n TRTM python=3.8
conda activate TRTM
Install python packages:
bash env_install.sh
TRTM dataset structure:
datasets
└── cloth_name
├── real
| ├── test/train
| | ├── {:06d}.real_depth.png
| └── └── {:06d}.real_color.png
| └── └── {:06d}.real_keypoint.txt
├── simu
| ├── test/train/val
| | ├── {:06d}.simu_depth.png
└── └── └── {:06d}.simu_mesh.txt
Download TRTM dataset from here and place it into 'TRTM/datasets'.
Download TRTM checkpoint from here and place it into 'TRTM/checkpoints'.
Train Cloth-GNN Model on template_square, use_simu only, and save to 'checkpoints/exp_train_half':
python main.py --phase train --cloth_name template_square --use_simu True
Test Cloth-GNN Model on template_square, use_real only, and from 'checkpoints/checkpoint_large_size.pt':
python main.py --phase test --cloth_name template_square --use_real True
Test Cloth-GNN Model on other_clothes named as smaller_square, larger_square, rectangle, or shirt:
python main.py --phase test --cloth_name other_clothes --use_real True
Step1: centralize and normalize your cloth depth images according to the following configurations (see datasets):
image_size: 720x720, flat_cloth_size: 480x480, flat_pixel_value: 190~195, 1cm_depth_pixel_value: 20.
Step2: create your_cloth_name and place your depth images into 'TRTM/datasets/your_cloth_name/real/test/'.
Step3: test Cloth-GNN Model on your cloth images using 'checkpoints/checkpoint_large_size.pt'.
python main.py --phase test --cloth_name your_cloth_name --use_real True
You can modify the blender files to generate new template mesh and training data.
Convolutional Mesh Regression for Single-Image Human Shape Reconstruction: https://arxiv.org/abs/1905.03244.
LEARNING MESH-BASED SIMULATION WITH GRAPH NETWORKS: https://arxiv.org/pdf/2010.03409.pdf
Implementation of meshgraphnets: https://github.com/wwMark/meshgraphnets
Li Gen's semester project: https://github.com/ligengen/TRTM
@article{wang2023trtm,
title={TRTM: Template-based Reconstruction and Target-oriented Manipulation of Crumpled Cloths},
authors={Wenbo Wang and Gen Li and Miguel Zamora and Stelian Coros},
booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
year={2024}
}
