A simple multi-3DCNN framework for action recognition using global and local information
python2.7 opencv3.0 keras2.0.8 tensorflow1.4
Use SSD (https://github.com/rykov8/ssd_keras) to detect people and save images.
KTH dataset:http://www.nada.kth.se/cvap/actions/
define the models
train the 3DCNN models with global and local information respectively.
extract 3dcnn feature with pretrained models and evaluate features with svm.
see details in model.py, train.py and evaluate.py scripts.
only use global information (total acc:81.40%)
only use local information (total acc:79.07%)
merge global and local information (total acc:87.44%)
