This is the project of my second internship, at the Computer Vision Laboratory of the University of Chile. In this work, I wrote the code to use the system of action classification developed by Georgia Gkioxari et al. in the paper "R-CNNs for Pose Estimation and Action Detection"

The system of detection is a convolutional neural network (CNN) of 8 layers. The first 5 layers are a group of convolutional layers, then layers 6,7 and 8 are fully connected (FC). After the last layer, there is a softmax function to transform the output in probabilities.

All the convolutional layers have a ReLU activation function, and the FC6-7 consider dropout. The next image shows the architecture of the implemented system for action detection.

The system was training and testing in PASCAL VOC 2012 action dataset. There are in total 10 different actions: jumping, phoning, playing instruments, reading, riding bike, riding horse, running, taking photos, using the computer and walking. The next images are a sample of the cited dataset