Exploration and visualization of complex data has become an integral part of life. But there is a semantic gap between the users and the visualization scientists. The priority of the users is usability while that of the scientists is techniques. Information-Assisted Visualization (IAV) can help bridge this gap, where additional information extracted from the raw data is presented to the user in an easily interpretable way. This thesis proposes some novel machine intelligence based systems for intuitive IAV.
The majority of the thesis focuses on Direct Volume Rendering, where Transfer Functions (TF) are used to color the volume data to expose structures. Existing TF design methods require manipulating complex widgets, which may be difficult for the user. We propose two novel approaches towards TF design.
In the data-centric approach, we generate an organized representation of the data through clustering and provide the user with some intuitive control over the output in the cluster domain. We use Spherical Self-Organizing Maps (SS)M) as the core of this approach. Instead of manipulating complex widgets, the user interacts with the simple SSOM color-coded lattice to design the TF.
In the image-centric approach, the user interaction with the data is direct and minimal. The user interactions create the training data, and supervised classification is used to generate the TF. First, we propose novel supervised classifiers that combine the local information available through Support Vector Machine-based classifiers and the global information available through Nonparametric Discriminant Analysis-based classifiers. Using these classifiers, we propose a
TF design method where the user interacts with the volume slices directly to generate the output.
Finally, we explore the use of IAV for home-based physical rehabilitation. We propose an information-assisted visual valuation framework which can compare a user’s performance of a physical exercise with that of an expert using our novel Incremental Dynamic Time Warping method and communicate the results visually through our color-mapped skeleton silhouette. All the proposed techniques are accompanied by detailed experimental results comparing them against the state-of-the-art. The results shows the potential of using machine learning techniques to achieve visualization tasks in a simpler yet more effective way.
Volume rendering using grid computing for large-scale volume data