The CVD Video Database

The CVD video database is developed to provide an useful tool for researchers in the validation and developing processes of no-reference (NR) objective video quality assessment (VQA) algorithms. It consists of 234 videos from five different scenes captured by 78 different cameras (mobile phones, compact camera, video camera, SLR). The subjective experiments are conducted following the Single-Stimulus (SS) procedure to collect ratings of video quality.

Moreover, objective test chart data are measured and can be used to analyze performance of capturing devices.

If you use this database in your research, we kindly ask that you follow the copyright notice and cite the following paper:

Video contents

Representative frames from different video scenes

Check example video clips here (small file size versions from the originals):

Subjective experiment description

Setup

We implement our experiments according to the Single Stimulus methodology using VQone MATLAB toolbox on high quality monitors (Eizo ColorEdge CG241W) with 1920x1200 pixel resolution in a dark room (ambient light < 20 lux). Display characterization measurement results can be found from here. Video stimuli were displayed at their original size of VGA (640 x 480) or HD (1280 x 720). The subjects viewing distance (80 cm) was controlled by a string hanging from a ceiling and they were instructed to keep their head steady next to it. The monitors were calibrated to according to sRGB (target values were: 6500 K, 80 lux, and gamma 2.2) using EyeOne Pro calibrator (X-rite co.). The laboratory setup is showed in the figure below.

Subjects

Subjects (n = 30, 30, 28, 33, 30, 32 and 27 for Tests 1 - 7 respectively) were naïve in a sense that they did not study or work with image quality or related fields. They were recruited through student mailing lists consisting mainly humanities and behavioral science students. Subjects’ vision was controlled for the near visual acuity, near contrast vision (near F.A.C.T.) and color vision (Farnsworth D15) before the participation. They received movie tickets as a reward.

Procedure

Subjects evaluated one video sample at a time and all video samples of one scene were presented in a row. The order of video samples and scenes was randomized. Subjects had the option to view video samples again as many times as they wanted.

Data

The results are processed and reported in the form of Mean Opinion Score (MOS) for the tested video samples. In addition, we provide the whole raw data from the subjective experiments instead of just pre-calculated mean opinion scores from each video sample. This allows further analyses to be made by those who wish to use this database and gives them better opportunity to utilize the data to its full potential.

Realignment study (test 7) contains the data from the additional study in which the mappings from the test and scene specific quality scales (test 1-6) to the global quality scale were formed. The global scale is valuable when studying and developing VQA algorithms. With the global scale, all of the samples (234 video samples in the case of the CVD2014) are in the same scale, and the performance analysis for algorithms can be conducted with a high number of samples. More details of the realignment study can be found here.

Download

Database is password protected. If you are interesting to download the database, please contact Toni Virtanen (firstname.lastname ( at ) helsinki.fi) or Mikko Nuutinen, (firstname.s.lastname ( at ) gmail.com).


Last modified: 2015/11/05