------------------------------------------------------------------------- LocVisVC 2019: Local visibility maps of artifacts and distortions in images under varying viewing conditions ------------------------------------------------------------------------- This dataset is composed of LocVis 2018 dataset and newly-collected measurements under varying viewing conditions: two display peak luminance levels and two viewing distances (angular resolutions in pixels per degree). This repository also contains the model checkpoint needed to run the Deep Photometric Visibility Metric (DPVM): https://github.com/ynyCL/DPVM This dataset was collected as a part of the project reported in the paper: Predicting visible image differences under varying display brightness and viewing distance Nanyang Ye, Krzysztof Wolski, Rafał K. Mantiuk. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019 Organization of the dataset --------------------------- The dataset is split into directories according to the type of introduced distortion. The new set with varying viewing conditions can be found in the directory "marking_view_cond". Each directory contains: * "marking_stimuli.csv" file with the list of all images. The file has the following columns scene - ID of a scene (content). Note that the scenes measured at multiple distortion levels will share scene IDs. level - ID of the distortion level base_fname - the common base file name of image files associated with this measured condition dataset - the name of the dataset n - how many observers marked each image peak_lum - peak luminance of the display used in the experiment ppd - angular resolution in pixels per visual degree * "_reference.png" - reference (non-distorted) images * "_test.png" - test (distorted) images * "_marking.png" - user marking. To get actual number of observers marking a part of a scene, the values should be computed as "round( M*n/255 )" where M is the value from the image, n is the number of observers (from "marking_stimuli.csv") and round is the rounding operator. * "P_obs.csv" - the look-up table with the P(p_att=p) probability distribution function (Fig. 6 in the paper). The value contain likelihood that the probability of attending a certain image region is equal p - it is p_att expressed as a random variable. * "likelihood.csv" - contains likelihood of P_det values The first row contains p_det values ranging from 0 to 1 The consecutive n+1 rows contain log10 likelihood values (note that those can be >0 because we do not normalize them). The first row is for the case when 0 observers mark the region, the second row for the case when 1 (out of n) observers mark the region, etc. The root directory also contains the names of the conditions that were used for each of the 5-fold data split into training and test sets: fold__test.txt is the test set for fold . The training set for the given fold combines the testing sets for all other folds. Copyright --------- Refer to the copyright.txt for the copyright information on images used in the dataset. Questions --------- Any questions regarding the dataset should be directed to: Rafal Mantiuk and Krzysztof Wolski (include both authors in the correspondence).