Action Recognition

This repo will host a collection of code examples and resources for using various methods to recognize human actions in videos.

I3D (Inflated 3D ConvNet)

I3D architecture was developed by the researchers at DeepMind for video classification. They have also released the Kinetics dataset which contains 600 different human actions with at least 600 video clips for each class.

To use i3d trained on Kinetics-600 dataset, clone this repo and follow the below commands.

git clone https://github.com/arunponnusamy/action-recognition.git
cd action-recognition/i3d/
python3 i3d_kinetics_600_tf_hub.py --label label_map_600.txt --video v_CricketBowling_g05_c02.avi 

You should get the below result or similar

playing cricket 0.8835184
brushing teeth 0.020662978
hammer throw 0.009957399
hopscotch 0.009599756
robot dancing 0.0078740865

You can also try with a different video clip containing any activity from the 600 classes listed in the label map.

Dependencies

• tensorflow 2.x
• tensorflow_hub
• opencv-python
• numpy