Realtime-Action-Recognition

Real-time Action Recognition using 3D CNNs for computationally limited platforms. The model developed is based upon MobileNetV2 and is able to perform human action recognition in real-time without GPU requirement. For model deployment to android app follow this link.

Requirements

• Pytorch
• Python 3.7+
• OpenCV

Datasets

• HMDB-51
• UCF-101

Place the datasets corresponding to path specified in datasets/dataset.py

Results

Model

2.4M parameters, 440M FLOPS, 10MB size

Accuracy

• HMDB-51: 51.3%
• UCF-101: 80.6%

Acknowledgements

• Thanks to okankop for arXiv:1904.02422 on which this project is build upon.