Dataset extraction for CNN model

Looking to create a CNN model that estimates a user's eye gaze direction of the quadrants forward, up, down, left and right. However, the dataset I want to train on needs to be extracted from the main dataset. The dataset being used is the MPII gaze dataset. There is an annotation.txt file that would allow for the specific images I want to train on to be extracted, but I don't have the knowledge to process it. The task is to write a script that can preprocess all the data extracting the 5 quadrants using the annotations file accurately.