Training Detail

Training configs are explained in lib/core/config.py. Different configs for different datasets and estimators are in config folder.

Dataset	Pose Estimator	3D Pose	2D Pose	SMPL
Sub-JHMDB	SimplePose		config
3DPW	EFT	config		config
3DPW	PARE	config		config
3DPW	SPIN	config		config
Human3.6M	FCN	config
AIST++	SPIN	config		config

Training Commands

You can directly train the model in different datasets and estimator settings using following commands

2D Pose

Sub-JHMDB Simplepose

python train.py --cfg configs/config_jhmdb_simplepose_2D.yaml --dataset_name jhmdb --estimator simplepose --body_representation 2D --sample_interval 10

3D Pose

3DPW SPIN

python train.py --cfg configs/config_pw3d_spin_3D.yaml --dataset_name pw3d --estimator spin --body_representation 3D --sample_interval 10

3DPW EFT

python train.py --cfg configs/config_pw3d_eft_3D.yaml --dataset_name pw3d --estimator eft --body_representation 3D --sample_interval 10

3DPW PARE

python train.py --cfg configs/config_pw3d_pare_3D.yaml --dataset_name pw3d --estimator pare --body_representation 3D --sample_interval 10

AIST++ SPIN

python train.py --cfg configs/config_aist_spin_3D.yaml --dataset_name aist --estimator spin --body_representation 3D --sample_interval 10

Human3.6M FCN

python train.py --cfg configs/config_h36m_fcn_3D.yaml --dataset_name h36m --estimator fcn --body_representation 3D --sample_interval 10

SMPL

3DPW SPIN

python train.py --cfg configs/config_pw3d_spin_smpl.yaml --dataset_name pw3d --estimator spin --body_representation smpl --sample_interval 10

3DPW EFT

python train.py --cfg configs/config_pw3d_eft_smpl.yaml --dataset_name pw3d --estimator eft --body_representation smpl --sample_interval 10

3DPW PARE

python train.py --cfg configs/config_pw3d_pare_smpl.yaml --dataset_name pw3d --estimator pare --body_representation smpl --sample_interval 10

AIST++ SPIN

python train.py --cfg configs/config_aist_spin_smpl.yaml --dataset_name aist --estimator spin --body_representation smpl --sample_interval 10

Useful configs

• Set cfg.TRAIN.RESUME = [checkpoint path], then you can resume training

• Set cfg.EXP_NAME = [your experiment name], then all the results would save in folder[time]_[your experiment name]