Introduction

PIFuHD was released from Facebook Research! I decided to try PIFuHD as an extension of playing with PIFu before.

In this article, we will describe the procedure from environment construction to executing the sample prepared on Windows.

The great PIFuHD repository is here [https://github.com/facebookresearch/pifuhd)

Execution environment

Windows10 64bit Home
Anaconda3
RTX2080Ti
CUDA 10.1

Environment

Creating a virtual environment

(base)$conda create -n pifu python=3.7.0
(base)$conda activate pifu

Installing the library (I have used pip, so please refer to this and install it with conda install ...)

(pifu)$conda install pytorch torchvision cudatoolkit=10.1 -c pytorch
(pifu)$pip install pillow==6.0.0 scikit-image tqdm opencv-python trimesh
(pifu)$conda install pyopengl

Install GLUT on Windows

This time, I installed Free GLUT instead of the original GLUT. Download "freeglut-MSVC-3.0.0-2.mp.zip" from here → unzip it I copied freeglut.dll located in freeglut \ bin \ x64 \ into C: \ Windows \ System32.

https://tadaoyamaoka.hatenablog.com/entry/2017/02/28/064625

I referred to this article. (Thank you!)

Install ffmpeg

https://fukatsu.tech/windows-ffmpeg

I used the procedure in this article as it is. (Thank you!)

Execution procedure

Code download

(pifu)$git clone https://github.com/facebookresearch/pifuhd.git
(pifu)$cd pifuhd

Model download

(pifu)$mkdir checkpoints
(pifu)$cd checkpoints
(pifu)$wget "https://dl.fbaipublicfiles.com/pifuhd/checkpoints/pifuhd.pt" -O pifuhd.pt
(pifu)$cd ..

If you don't have wget, download pifuhd.pt directly from here and store it under checkpoints.

Fix code

I got an error when using OpenGL in my environment, so I modified lib / render / gl / render.py as follows.

[render.py]
class Render:
    def __init__(self, width=1600, height=1200, name='GL Renderer',
                 program_files=['simple.fs', 'simple.vs'], color_size=1, ms_rate=1):
        self.width = width
        self.height = height
        self.name = name
        self.display_mode = GLUT_DOUBLE | GLUT_RGB | GLUT_DEPTH
        self.use_inverse_depth = False

        global _glut_window
        if _glut_window is None:
            glutInit()
            glutInitDisplayMode(self.display_mode)
            glutInitWindowSize(self.width, self.height)
            glutInitWindowPosition(0, 0)
            #Fix here
            # _glut_window = glutCreateWindow("My Render.")
            _glut_window = glutCreateWindow(b"My Render.")

Execute the command of demo.sh

(pifu)$python -m apps.simple_test
(pifu)$python -m apps.render_turntable -f ./results/pifuhd_final/recon -ww 512 -hh 512

When you execute the above command, you will see the following animation on the screen!

In addition, I think that "result_test_512.mp4" is also output under results / pifuhd_final / recon.

in conclusion

This time, I tried running the PIFuHD sample published by Facebook Research on Windows. Basically, if you check README.md, you can easily try it.

In the future, I will try using my own images instead of the prepared samples. PIFu had to be Sementic-Segmentation, but in PIFuHD it seems that the key point information obtained by Pose-Estimation is sufficient. Officially, a trial method using OpenPose was published, but I think that anything can be done if the output format can be matched by the Pose-Estimation algorithm separately.

I tried running PIFuHD on Windows for the time being