Registration deadline in ZeUS: Oct 15 - Nov 15, 2024
Due date projects: Jan 15, 2025
Next opportunity to register after the current deadline: 15.01.-15.02.2025.
When interested in one of the projects, contact us asap(e.g., dietmar.saupe@uni-konstanz.de), and we will invite you for a meeting to talk and discuss the project/thesis prospects.
These projects and the following theses can be scaled such that they are suitable for either Bachelor or Master level.
We do not offer accompanying seminars. Instead, we arrange and advise a topic related to a chosen project to be presented in an existing seminar in another research group.
Image compression technology can encode images at much lower bitrates than that of the raw format, i.e., 24 bits/pixel for RGB color images. Yet, the visual image quality of the decoded image can be so good that is very hard to see any differences to the original source image. When the bitrate gets lower, distortions become more visible and at the end even annoying. The degree of perceived distortion can be measured in experiments with human observers and expressed in so-called JND units. This unit is defined as the perceived distance of the distorted image from the source such that the probability that a random observer can correctly identify the distorted image is 75%. Note that the chance rate for the correct response is already 50%. Now, here is the challenge: Given a source image and a number of encoded images, can you estimate their perceived quality in JND units?
The standardization project JPEG AIC is providing fresh data for this from experiments with 5 images, 5 compression engines operating at 10 different bitrates each, thus, 250 cases(Testolina 2024). It will be your job to design, implement and evaluate algorithms for this challenging research question, with our help, of course. In the project it suffices to provide a proof-of-concept using an elementary algorithm. Of course, there are already methods(and code) available for other types of IQA that can be adapted to the problem on hand and we also have a proposal.
JPEG AIC Test images: Compression artifacts generated with JPEG, JPEG 2000, HEVC Intra, VVC Intra, and JPEG XL at multiple quality levels(0.25 JND spacing)
The JPEG AIC quality assessment used a boosting technique to make subtle differences between a high quakity decoded image and the source image visible.
In January 2025, JPEG plans to publish a Draft Call-for-Contributions for solutions to this research question. At that time an extension of the dataset will be provided for training. This aligns very well with the time allocated for the thesis. Therefore, there is a chance at the end to submit your solution to the JPEG Committee for consideration in context of a future ISO standard for full-reference objective image quality assessment.
Literature
Michela Testolina, Mohsen Jenadeleh, Shima Mohammadi, Shaolin Su, João Ascenso, Touradj Ebrahimi, Jon Sneyers, and Dietmar Saupe, Fine-grained subjective visual quality assessment for high-fidelity compressed images, arXive, Oct 2024.
A practical guide and software for analysing pairwise comparison experiments
Project: Ordering bias in two-alternative forced choice experiments
Advisor: Prof. Dietmar Saupe
This is a topic originally coming from psychophysics/psychometrics requiring algorithms for data analysis, yet with implications for testing procedures in multimedia research.
Wikipedia explains:
Psychophysics quantitatively investigates the relationship between physical stimuli and the sensations and perceptions they produce. Psychophysics has been described as"the scientific study of the relation between stimulus and sensation" or, more completely, as"the analysis of perceptual processes by studying the effect on a subject's experience or behaviour of systematically varying the properties of a stimulus along one or more physical dimensions".
In digital signal processing, insights from psychophysics have guided the development of models and methods for lossy compression. Psychometrics is a field of study within psychology concerned with the theory and technique of measurement. It is concerned with the objective measurement of latent constructs that cannot be directly observed. The levels of individuals on nonobservable latent variables are inferred through mathematical modeling based on what is observed from individuals' responses to items on tests and scales.
Biases can occur in many ways in psychological surveys, for example:
In visual quality assessment studies a similar order bias has been observed: We have carried out a number of large-scale crowdsourcing studies using pair comparisons collecting several hundred thousand binary and ternary responses(Jenadeleh 2023, Testolina 2024). In this project we will analyze these datasets with respect to order bias. Order bias is a term to define a condition in which the order of your questions and answer options can affect how respondents give feedback. For example, in a binary choice question, respondents may be more likely to choose the first option over the second.
There are several methods to handle such biases in order to model bias and to reduce its detrimental effect on the outcome of the experiment. García-Pérez 2017 presents one and Ellinghaus 2024 gives references to three more.
We will first analyze our datasets to quantify the order bias in them.
Then we will study and implement some of debiasing methods and
Check to what extent these provide statistically better models for the empiricaldata.
Available projects in Visual Computing
Project: Full-reference objective image quality assessment (FR-IQA)
Project: Ordering bias in two-alternative forced choice experiments