Decanus to Legatus: Synthetic training for 2D-3D human pose lifting

This is the original implementation of the ACCV 2022 Paper "Decanus to Legatus: Synthetic training for 2D-3D human pose lifting"

Required packages:

• pytorch
• Pillow (PIL)
• matplotlib
• numpy

Sample Code

• To reproduce a handcrafted distribution, run:

python train.py --task make-distribution

or

python train.py --task md

The reproduced distribution will be saved inside './distribution/handcrafted' folder.

• To draw generation samples from our handcrafted distribution, run:

python train.py --task generation

or

python train.py --task g

The samples will be saved inside './examples' folder.

• To train, run:

python train.py --task train

or

python train.py --task t

The weights will be saved in 'model' folder under name as 'lifter_(epoch).pth'

• To show inference examples of our pretrained model on several COCO dataset samples, run:

python train.py --task inference

or

python train.py --task i

The samples will be saved inside './examples' folder.

• To do all former 4 steps in the introduced order, run:

python train.py