- Partner: CSIR MIAS
- Students: Purity Molala (UL), Raesetje Sefala (Wits), Siyabonga Mbonambi (NMU)
- Project Lead: Dr. Michael Burke
- Year: 2017/2018
In the recent past, computer vision, which is an interdisciplinary field that deals with how algorithms can be designed to gain understanding from digital images or videos, has gained extensive attention. One of the main objectives in computer vision is explore a large data set of images with the aim to flag images that are useful to the domain expert (here a domain expert is someone that has knowledge about the images that they would generally want to see in the data set, or specific images that are of interest to them). As a result, the ability to predict the importance of images is highly desirable in computer vision.
However, computer vision requires a great deal of labelled data to be available and asking a domain expert to go through the manual process of labelling data is a daunting and inefficient task. To overcome this limitation, we made use of a ranking algorithm that uses pairwise image comparisons. Here, a user is presented with two images, and asked which they prefer. These pairwise comparisons are used to infer the interest value of images being compared. Unfortunately, this process is time consuming,. so the ability to speed up this process is the primary goal of this project. In this work, we are developing a software tool that allows users to explore a data set of images with limited labelling effort, using ranking algorithms to flag images of interest and highlight content within these that makes the images interesting, as shown in the image below. This tool will give domain experts the ability to rapidly explore their data, and build models that are able to highlight image content of interest in the process. In addition, an active sampling strategy has been developed to reduce the time required to analyse a data set, thereby providing valuable, automatic insights to the domain expert with significantly reduced effort.