The NTURGB+D 2D detection results are provided by pyskl using HRNet.
- Download
ntu60_hrnet.pkl
andntu120_hrnet.pkl
todata/action/
. - Download the 1-shot split here and put it to
data/action/
.
Train from scratch:
# Cross-subject
python train_action.py \
--config configs/action/MB_train_NTU60_xsub.yaml \
--checkpoint checkpoint/action/MB_train_NTU60_xsub
# Cross-view
python train_action.py \
--config configs/action/MB_train_NTU60_xview.yaml \
--checkpoint checkpoint/action/MB_train_NTU60_xview
Finetune from pretrained MotionBERT:
# Cross-subject
python train_action.py \
--config configs/action/MB_ft_NTU60_xsub.yaml \
--pretrained checkpoint/pretrain/MB_release \
--checkpoint checkpoint/action/FT_MB_release_MB_ft_NTU60_xsub
# Cross-view
python train_action.py \
--config configs/action/MB_ft_NTU60_xview.yaml \
--pretrained checkpoint/pretrain/MB_release \
--checkpoint checkpoint/action/FT_MB_release_MB_ft_NTU60_xview
Evaluate:
# Cross-subject
python train_action.py \
--config configs/action/MB_train_NTU60_xsub.yaml \
--evaluate checkpoint/action/MB_train_NTU60_xsub/best_epoch.bin
# Cross-view
python train_action.py \
--config configs/action/MB_train_NTU60_xview.yaml \
--evaluate checkpoint/action/MB_train_NTU60_xview/best_epoch.bin
Train from scratch:
python train_action_1shot.py \
--config configs/action/MB_train_NTU120_oneshot.yaml \
--checkpoint checkpoint/action/MB_train_NTU120_oneshot
Finetune from a pretrained model:
python train_action_1shot.py \
--config configs/action/MB_ft_NTU120_oneshot.yaml \
--pretrained checkpoint/pretrain/MB_release \
--checkpoint checkpoint/action/FT_MB_release_MB_ft_NTU120_oneshot
Evaluate:
python train_action_1shot.py \
--config configs/action/MB_train_NTU120_oneshot.yaml \
--evaluate checkpoint/action/MB_train_NTU120_oneshot/best_epoch.bin