TITLE: Data Analytics For Image Visual Complexity and Kinect-Based Videos of Rehabilitation Exercises
PRESENTER: Elham Saree
DATE: May 6, 2019
LOCATION: Math & Computer Science, room 148
CHAIR: Prof. Ioannis Paschalidis
COMMITTEE:Prof. Margrit Betke (Advisor), Prof. Prakash Ishwar, Prof. Brian Kulis, Prof. Terry Ellis
With the recent advances in computer vision and pattern recognition, methods from these fields are successfully applied to solve problems in various domains, including health care and social sciences. In this thesis, two such problems, from different domains, are discussed. First, an application of computer vision and broader pattern recognition in physical therapy is presented. Home-based physical therapy is an essential part of the recovery process in which the patient is prescribed specific exercises in order to improve symptoms and daily functioning of the body. However, poor adherence to the prescribed exercises is a common problem. In our work, we explore methods for improving home-based physical therapy experience. We begin by proposing DyAd, a dynamically difficulty adjustment system which captures the trajectory of the hand movement, evaluates the user's performance quantitatively and adjusts the difficulty level for the next trial of the exercise based on the performance measurements. Next, we introduce ExerciseCheck, a remote monitoring and evaluation platform for home-based physical therapy. ExerciseCheck is capable of capturing exercise information, evaluating the performance, providing therapeutic feedback to the patient and the therapist, checking the progress of the user over the course of the physical therapy, and supporting the patient throughout this period. In our experiments, Parkinson patients have tested our system at a clinic and in their homes during their physical therapy period. Our results suggests that ExerciseCheck is a user-friendly application and can assist patients by providing motivation, and guidance to ensure correct execution of the required exercises.
As the second application, and within computer vision paradigm, we focus on visual complexity, an image attribute that humans can subjectively evaluate based on the level of details in the image. Visual complexity has been studied in psychophysics, cognitive science, and, more recently, computer vision, for the purposes of product design, web design, advertising, etc. We first introduce a diverse visual complexity dataset which compromises of seven image categories. We collect the ground-truth scores by comparing the pairwise relationship of images and then convert the pairwise scores to absolute scores using mathematical methods. Furthermore, we propose a method to measure the visual complexity that uses unsupervised information extraction from intermediate convolutional layers of deep neural networks. We derive an activation energy metric that combines convolutional layer activations to quantify visual complexity. The high correlations between ground-truth labels and computed energy scores in our experiments show superiority of our method compared to the previous works. Finally, as an example of the relationship between visual complexity and other image attributes, we demonstrate that, within the context of a category, visually more complex images are more memorable to human observers.