gengshan-y / dyn_pose

Dynamic pose estimation in MXNet. faster-rcnn detection + cpm pose estimation + LSTM dynamical modeling

dyn_pose

Training and inference code for LSTM-based action recognition model. Online version to be released later.

Steps

• models
• pre-process data: preprocess.py / split.py
• extract pose: extract.py
• train model: train_lstm.py
• evaluate model: infer_lstm.py
• offline demo: example.py

Notes

• modify label_num in model/lstm.config

Utils

• display.ipynb: display pose estimation results
• run.sh: evaluation script
• test.ipynb: miscellaneous

TODO

• reduce spatial net size

Reference

• Convolutional Pose Machines