Awesome

TailorNet Dataset

This repository is a toolbox to process, visualize the dataset for "TailorNet: Predicting Clothing in 3D as a Function of Human Pose, Shape and Garment Style" (CVPR 2020 Oral)

[model repository][arxiv][project website][YouTube]

Update

2021/2/2 dataset uploaded to Baidu Drive
2021/1/7 data generation codes
2020/12/7 short pants, skirt are available
2020/7/31 pants, shirt are available

Requirements

python3
pytorch
chumpy
opencv-python
cython

SMPL model

1. Register and download SMPL models here
2. Unzip SMPL_python_v.1.0.0.zip and put smpl/models/*.pkl in ROOT/smpl(specify ROOT in global_var.py)
3. Run smpl_lib/convert_smpl_models.py

Data preparation

All data is available in the following links:
Data or
Baidu Drive(password：TLNT)

1. Download meta data (dataset_meta.zip) of the dataset

2. Download one or more sub-dataset (other garment classes are coming soon)
   t-shirt_female(6.9G)
   t-shirt_male(7.2G)
   old-t-shirt_female(10G)
   t-shirt_female_sample(19M)
   shirt_female(12.7G)
   shirt_male(13.5G)
   pant_female(3.3G)
   pant_male(3.4G)
   short-pant_female(1.9G)
   short-pant_male(2G)
   skirt_female(5G)

3. Specify the variable ROOT in global_var.py

4. Unzip all downloaded files to ROOT

Dataset Description

Currently, we have 6 garment classes (t-shirt, shirt, pant, skirt, short-pant, old-t-shirt). In TailorNet paper, we trained and tested our model using old-t-shirt. Compared to old-t-shirt, t-shirt has a different topology, higher quality and larger style variation. Use old-t-shirt if you want a fair comparison with the results in our paper.

The dataset structure looks like this:

ROOT
----smpl
----apose.npy
----garment_class_info.pkl
----split_static_pose_shape.npz

----<garment_class>_<gender> (e.g., t-shirt_female)
--------pose/
------------<shape_idx>_<style_idx> (e.g., 000_023)
--------shape/
--------style/
--------style_shape/
--------avail.txt
--------pivots.txt
--------test.txt
--------style_model.npz

We provide apose.npy, garment_class_info.pkl and split_static_pose_shape.npz separately in dataset_meta.zip, and each <garment_class>_<gender> in a separate zip file.

• split_static_pose_shape.npz contains a dictionary {'train': <train_idx>, 'test': <test_idx>} where <train_idx> and <test_idx> are np arrays specifying the indices of poses which goes into train and test set respectively.

• garment_class_info.pkl contains a dictionary {<garment_class>: {'f': <f>, 'vert_indices': <vert_indices>} } where <vert_indices> denotes the vertex indices of high resolution SMPL body template which defines the garment topology of <garment_class>, and <f> denotes the faces of template garment mesh.

• apose.npy contains the thetas for A-pose on which garment style space is modeled.

• For each <garment_class>_<gender>,

  • shape directory contains uniformally chosen shape(beta) parameters.

  • style_model.npz contains a dictionary with these variables: pca_w, mean, coeff_mean, coeff_range. For given style gamma, garment vertices can be obtained using the following equation:

    • pca_w * (gamma + coeff_mean) + mean

  • style directory contains uniformally chosen style(gamma) parameters.

  • All styles are simulated on all shapes in A-pose and results are stored in style_shape directory. Out of those, shape_style pairs (also called pivots) with feasible simulation results are listed in avail.txt.

  • pivots.txt lists those pivots which are chosen as per the algorithm described in subsection - Choosing K Style-Shape Prototypes - to simulate training data. test.txt lists additional pivots chosen to generate testing data.

  • Each chosen pivot, denoted as <shape_idx>_<style_idx>, is simulated in few pose sequences. Simulation results are stored in pose/<shape_idx>_<style_idx> directory as unposed garment displacements. (Garment displacements are added on unposed template before applying standard SMPL skinning to get the final garment. See paper for details.)

  • pose/<shape_idx>_<style_idx> also contains displacements for smoothed unposed garment.

Usage

If you want to convert the data to the mesh format (e.g., .obj), please check Line 44 to Line 64 in visualize_dataset.py. This code converts the TailorNet data sequences into meshes (gar_v, gar_f are vertices and faces of the garment) and renders them.

Visualize the dataset

1. Install the renderer

cd utils/render_lib
python setup.py build_ext -i

2. Run the visualizer

python visualize_dataset.py

Dataset Generation

Download datagen_assets.zip and unzip it to ROOT. Please check readme.md in each directory for detail

1. style_pca
   Scripts that process garment registrations and model the garment style space.
2. simulation_style
   Simulate all (style, shape) combinations in A-pose.
3. pivots
   Generate pivots and test set.
4. simulation_pose
   Simulate different poses for pivots and test (style, shape).

Since our raw data is not public and simulation in Marvelous Designer cannot be scripted, these codes are only for reference. If you want to simulate your own data, make sure you understand most code and the paper, so that you can modify parameters that are highly dependent of the data.

Here is an example video of simulation in Marvelous Designer.

Count the dataset

python count_data.py
——————————————————————————————————————————————————————————————————————————————
|          |          |    train style_shape|     test style_shape|          |
|     class|    gender|train pose| test pose|train pose| test pose|     total|
——————————————————————————————————————————————————————————————————————————————
|   t-shirt|    female|     14589|      3309|       776|       224|     18898|
|   t-shirt|      male|     14397|      3353|       815|       185|     18750|
|     shirt|    female|     14553|      3342|       856|       144|     18895|
|     shirt|      male|     14322|      3328|       831|       169|     18650|
|      pant|    female|     14569|      3430|       805|       195|     18999|
|      pant|      male|     14562|      3423|       793|       203|     18981|
|short-pant|    female|     14546|      3451|       804|       196|     18997|
|short-pant|      male|     14563|      3426|       796|       203|     18988|
|     skirt|    female|     14554|      3444|       803|       197|     18998|
|                total|    130655|     30506|      7279|      1716|    170156|
——————————————————————————————————————————————————————————————————————————————

TODO

• Dataset generation codes
• Style space visualizer
• Blender visualizer
• Shirt, pants, skirt
• T-shirt
• Basic visualizer

Citation

Cite us:

@inproceedings{patel20tailornet,
        title = {TailorNet: Predicting Clothing in 3D as a Function of Human Pose, Shape and Garment Style},
        author = {Patel, Chaitanya and Liao, Zhouyingcheng and Pons-Moll, Gerard},
        booktitle = {{IEEE} Conference on Computer Vision and Pattern Recognition (CVPR)},
        month = {jun},
        organization = {{IEEE}},
        year = {2020},
    }