This repository contains the trained neural network (TensorFlow) models, associated with the paper "Lightweight Multitask Learning for Robust JND Prediction using Latent Space and Reconstructed Frames", published in IEEE T-CSVT 2024. The code associated with these models can be accessed through the paper's GitHub page. Paper The paper is available on IEEEXplore and a preprint is available on TechArxiv. Requirements Tensorflow FFmpeg Dataset Our evaluation is conducted on VideoSet and MCL-JCI datasets. Usage Our pretrained models are capable of predicting JND values, and they can also be employed for training on a custom dataset. Note: The dataset used for training and testing should have such a structure. - rootdir/ - train/ - img#1 - ... - JND-Levels.txt (a file containing the 3 JND levels per image: first column for the first JND, second column for the second JND, and third column for the third JND level) - valid/ - img#1 - ... - JND-Levels.txt (a file containing the 3 JND levels per image: first column for the first JND, second column for the second JND, and third column for the third JND level) - test/ - img#1 - ... - jnd1train/ - img#1 - ... - jnd1valid/ - img#1 - ... - jnd2train/ - img#1 - ... - jnd2valid/ - img#1 - ... - jnd3train/ - img#1 - ... - jnd3valid/ - img#1 - ... Testing For prediction with LAT or REC model, the following commands can be used. python3 [LAT.py or REC.py] test --jnd_value [JND1 or JND2 or JND3] --data_dir "Path-to-the-rootdir/" --model_weights_path "Path-to-the-pretrained-model/" --result_path "Path-to-save-test-results/" --JND_Recon_Models_Path "Path-to-the-pretrained-JND-Reconstruction-models/" For prediction with E2E-LAT or E2E-REC model, the following commands can be used. python3 [E2ELAT.py or E2EREC.py] test --jnd_value [JND1 or JND2 or JND3] --data_dir "Path-to-the-rootdir/" --model_weights_path "Path-to-the-pretrained-model/" --result_path "Path-to-save-test-results/" --ImgReconstrution_Model_Path "Path-to-the-pretrained-Img-Reconstruction-models/" For prediction with MJ-LAT or MJ-REC model, the following commands can be used. python3 [MJLAT.py or MJREC.py] test --data_dir "Path-to-the-rootdir/" --model_weights_path "Path-to-the-pretrained-model/" --result_path "Path-to-save-test-results/" --JND_Recon_Models_Path "Path-to-the-pretrained-JND-Reconstruction-models/" More details about the associated codes can be found on the github page: https://github.com/sanaznami/MTL_JND