README.md

Youtubehand

Introduction

---

⚠Unofficial⚠ PyTorch implementation of Weakly-Supervised Mesh-Convolutional Hand Reconstruction in the Wild, CVPR 2020

Install

---

My codebase is developed based on Ubuntu 18.06 Python 3.7.13 CUDA 11.4.

• Install pytorch and torchvision according to your CUDA and python version.

• Requirements

  pip install -r requirements.txt
  

• Install MPI-IS Mesh

• You should accept MANO LICENCE. Download the MANO model and another files.

The resulting data structure should follow the hierarchy as below.

${REP_DIR}
|--conv
|--data
	|--freihand
|--datasets
|--images
|--out
	|checkpoints
	|board
	|demo
	|eval
|template
|utils
|...
|...

Trained Model Download

---

• The pre-trained HRNet can be downloaded according to METRO

• Download the trained model from GoogleDrive

• Here I report my re-produced results on FreiHAND

Methods	Backbone	PA-MPJPE	PA-MPVPE	#Params
Origin	ResNet50	8.4	8.6	-
Reproduced	ResNet18	8.6	8.7	36M
Reproduced	ResNet50	7.7	7.7	419M
Reproduced	HRNet-W64	7.2	7.4	519M

Demo

---

• Put the input images in the images folder
• Run

python main.py --split demo --resume --exp_name $exp_name under the out folder e.g. global-resnet18$

Train

---

• Follow METRO to download FreiHAND dataset.
• Run

# resnet18
python main.py --split train --batch_size 64 --epochs 38 --decay_step 30 --backbone resnet18 --out_channels 64 128 256 512 --exp_name global-resnet18

Evaluation

---

• Run

python main.py --split eval --exp_name $exp_name under the out folder e.g. global-resnet18$

Acknowledgement

---

The implementation modifies codes or draws inspiration from:

• HandMesh
• METRO
• manopth