Readme.txt~

dataset: sample training set from GTEA
 - adversial: each training data is 32 consecutive video frames 
 - adversial: corresponding gaussian masked fixation maps for supervised training


filelist: a list of filenames to read and test on

matlab: evluation code to display predicted saliency maps and evaluate its performance
 - run computeAUCAAEAdversaial_gtea.m

we extend our CVPR method by adding in a DFG-P pathway which predicts gaze prior maps based on the task at hand extracted from the current input frame
 - run computeAUCAAEAdversarial_gtea_fusion.m

models: to store saved models at checkpoints

OSdatasetProcess: matlab code for pre-processing Object Search Dataset (OS)
 - VXY: ground truth gaze data

results: to store predicted saliency maps in .mat converted from Torch

vis: to store generated future frames, backgrounds, masks

torchMM: 
 - main_GAN: training adversarial networks
 - main_gazePred: training gaze prediction module (cvpr; DFG-G pathway)
 - main_gazePrior: training gaze prior map generation module (DFG-P pathway)
 - generateGAN: testing file to generate future frames and store in vis folder
 - generateGaze: testing file to generate future gazes and store in .mat in results
 - generateGazePrior: testing file to generate future gaze priors and store in .mat in results

OSdatasetProcess:
 - GenerateFrameOSDataset: generate frames from raw vides
 - GenerateGazeOSDatast: generate gaze from raw data recorded from eyetrackers
 - GenerateAdversarialTrainingImage: combine frames into training data formats: consecutive 32 frames into 1 long frame
 - GenerateAdversarialTrainingMask: combine gaussian masked ground truth fixation maps: consecutive 32 fixation maps into 1 long frame

**The source code can be run directly. In order to train the network, pls download the GTEA, GTEAplus and OS datasets.
Use the OSdataset sample codes to pre-process the data to generate training and testing dataset.