human2b / temporal_inverse_kinematics

a learning-based human inverse kinemtics. A graph convolution network is constructed to predict SMPLx joint angles from a tepmoral sequence of relative 3d poses in COCO format.

Deep learning-based inverse kinematics

Introduction

This project presents a human inverse kinemtics solution based on deep learning. A graph convolution network is constructed to predict SMPLx joint angles from a tepmoral sequence of relative 3d poses in COCO format.

Install

conda env create -n motion python=3.8.2
conda activate motion
pip install -r requirements.txt

prepare the dataset

• register and download the Amass dataset, save it to the folder AMASS_DIR

• register and download the SMPLx models, save it also to the folder AMASS_DIR

Run test inference.

download the amass dataset, SMPLx models and save it to the folder AMASS_DIR

python inference.py ./data/sample_3d_poses/dance_contemporary.npz AMASS_DIR

Train the model

python ./model_wrap.py --amass AMASS_DIR  --data_dir DIR_TO_SAVE_MODELS --smpl_mean ./data/smpl/smpl_mean_params.npz