This repository contains a code for obtaining a 3D reconstuction of a garment. This is a "proof-of-concept" version of a potential solution for a garment reconstruction task by monocular images. Currently the code supports 2 types of garment ('shirt', 'pants').

For reconstruction you need to have an image of clothing lying on a flat surface. There are 2 variants of usage:

• using front side image of a garment;
• using both: front and back side images.

The reconstruction lies in transfering texture from the image to the corresponding 3D template-mesh of the garment type. The garment type needs to be specified with the input.

• OS: Ubuntu 20.04.
• conda
• cuda-toolkit

conda create -n grecon python=3.7.9
conda activate grecon

# For CUDA 10
# conda install -c pytorch -c conda-forge pytorch=1.10.1 torchvision=0.11.2 cudatoolkit=10.2
# For CUDA 11
conda install -c pytorch -c conda-forge pytorch=1.10.1 torchvision=0.11.2 cudatoolkit=11.1

pip install -r requirements.txt
mkdir input

cd kgdet/deepfashion2_api/PythonAPI
pip install -e .
cd ../../mmdetection
python setup.py develop
cd ../../

Download checkpoints and meshes and unzip them at the root folder.

Place a garment images at ./input folder. You can also use another folder for that purpose, but you will need to specify it for demo.py script via --root argument.

You can dowload demo-data here and extract it at the root folder.

Any of scripts needs to invoked from the root directory. For now there are available 2 possible classes for the reconstruction: shirt and pants.

Note, that a folder for output needs to be empty, otherwise its content will be removed automatically.

After invoking demo.py script check results at the folder specified via --out argument.

usage: demo.py [-h] [--root ROOT] [--out OUT] [--mesh MESH]
               [--checkpoints CHECKPOINTS] --type {pants,shirt} --front FRONT
               [--back BACK] [--verbose]

optional arguments:
  -h, --help            show this help message and exit
  --root ROOT           Directory, where images are placed.
  --out OUT             Directory for output results.
  --mesh MESH           Directory of a folder containing pattern-meshes.
  --checkpoints CHECKPOINTS
                        Directory of a folder containing checkpoints for
                        kgdet.
  --type {pants,shirt}  Target type of garment for reconstruction.
  --front FRONT         Path to image of a front side of a garment.
  --back BACK           Path to image of a front side of a garment.

python demo.py --root ./input --type shirt --front shirt6.jpg --back shirt6_b.jpg 
python demo.py --root ./input/pants1 --type pants --front 1.png
python demo.py --root ./input/pants3 --type pants --front 3_f.png --back 3_b.png