The projects envisions to use CNNs for visual saliency detection task
Short description of the codes::
a. map_create.m :: creates saliency maps on ImageNet using Graph Based Visual Saliency Model whose library is under the gbvs folder
b. saliency_generate.sh :: a wrapper to call map_create.m with proper arguements
c.startup.m :: file required to include the gbvs library path in Matlab's search path
d. imagenet_saliency_prep.py :: iterates over the entire ImageNet folders to generate a consolidates training set with ImageNet RGB images and corresponding saliency map
e.dataset_ground_withmat.py:: creates the ground truth training set using eye fixation maps of iSUN, SALICON and MIT datasets
f.final_attempt_pretrain.py:: A Lasagne-Nolearn variant of training CNN on ImageNet saliency maps for pre-training phase
g.final_attempt_posttrain.py:: Uses pre-learned weights to train a CNN model on actual human eye fixation ground truths
h.predict2gray.m : converts the final predicted saliency vectors to gray scale images
Some data processing parts have been coded in Matlab because still major competitions on saliency detections upload helper
functions in Matlab