Peter Remøy Paulsen has started his master thesis and joined the AdMiRe project on August 2nd. He is interested in all aspects of data science and image processing from software engineering, visualising data with python, using statistics and specially looking at aspects of quality using objective metrics and subjective tests. He is very fluent at all the techniques that involve quality assessment research and he has showed a strong interest in finding the right methods to answer his research question in the best possible way. He has shown a strong autonomy while at the same time willing to join energies with the direction of the AdMiRe project.
In his thesis he is looking at the correlation between the objective measures of pixel accuracy, intersection over union, dice coefficient and the subjective measures of satisfaction, the noticing of artifacts and the level of annoyance in a machine learning based foreground extraction algorithm which is applied to a number of video sequences.
In order to investigate the mentioned research question Peter has produced a number of video sequences challenging the algorithm with specific visual tests. The tests are designed to find out the strengths and weaknesses of the foreground extraction module. These are using various background images from NTNU (Figure1), different clothing, holding objects, and counting with the fingers (Figure 2). A total of 12 videos have been produced and have been evaluated with objective metrics and subjective tests.
The videos have been first recorded in a green screen and the background has been extracted. Later on, a background image has been inserted into the segmented video and the result has been fed into the machine learning based foreground extractor. The silhouettes of both, the ground truth which is the result of the chroma key and the foreground extractor module are compared. Then the deviation of pixels has been analysed and the numerical difference is compared to the results given by the subjective ratings. This study will help us understand the correlation between the scores produced by counting pixels and the scores given by the test subjects. At this point, everything seems to indicate that some videos have a correlation between the objective and subjective scores and some others diverge. If you want to know more, the results will be published in his master thesis very soon.
Figure 1 Figure 2
The Sense-IT team at NTNU is very happy having him on board and the AdMiRe team has greatly benefit from his work. Peter will finalise his thesis soon and we all want to help him in his last steps.
Go Peter!