This is my research in the project for Dynamic video retrieval. I try to use a sequence of action data (lie algebra data format) within a action label to predict the following frames.The lesser input frames, the longer prediction frames, and the lower MAE, the better.

Multi-scale Human action relation modeling Action encoding method Multi-scale spatio-temporal imfomation exchanging

python train.py --action walk --dataset Human

python train.py --dataset Human --training False --visualize False --action walk

you can select the actions in the folder './data/h3.6m/Train'

change the checkpoint_dir & input_data_dir in pre_one.py and run pre_one.py