Awesome

Human 2D images to 3D models

Project to put the Unite the People project in a Docker image. The project contains 2 docker images because the pose and segmentation scripts require different environments.

• up_caffe: run the pose prediction with the p91 model
• up_deeplab: run the segmentation script with the s31 model

Usage

When you run the containers you need to mount a local folder which contains the source images that you want to use for pose predictions and segmentations.

1. Download

$ docker pull lukin0110/up_caffe
$ docker pull lukin0110/up_deeplab

2. Execute pose prediction

$ docker run -it -v "$(pwd)"/input:/input lukin0110/up_caffe pose input/debruyne1.jpg

3. Execute bodyfit

$ docker run -it -v "$(pwd)"/input:/input lukin0110/up_caffe bodyfit input/debruyne1.jpg

4. Execute segmentation

It's required to use NVIDIA Docker to run the segmentation since CUDA is being used.

$ nvidia-docker run -it -v "$(pwd)"/input:/input lukin0110/up_deeplab segmentation input/debruyne1.jpg

Setup from source

Before you execute the script: download SMPL_python_v.1.0.0.zip from http://smpl.is.tue.mpg.de/ and put it in the ./models folder. You need an account to download the package.

./prepare.sh

Build image:

$ docker-compose build

Generate pose prediction:

$ docker-compose run caffe pose input/debruyne1.jpg

This will generate .npz and .png files in the input folder.

Generate 3D body:

$ docker-compose run caffe bodyfit input/debruyne1.jpg

Generate segmentation:

$ nvidia-docker-compose run deeplab segmentation input/debruyne1.jpg

Install nvidia docker compose if it is not present yet: pip install nvidia-docker-compose.

Push to docker hub

$ docker tag demo2d3d_caffe:latest lukin0110/up_caffe:latest
$ docker tag demo2d3d_deeplab:latest lukin0110/up_deeplab:latest
$ docker push lukin0110/up_caffe:latest
$ docker push lukin0110/up_deeplab:latest