Measuring the Detection of Objects under Simulated Visual Impairment in 3D Rendered Scenes

University of Minnesota Ph.D. dissertation. September 2018. Major: Psychology. Advisor: Daniel Kersten. 1 computer file (PDF); vii, 70 pages.A space is visually accessible when a person can use their vision to travel through the space and to pursue activities intended to be performed with vision within that space. Previous work has addressed the difficulty of evaluating the detection of objects in real spaces by observers with simulated visual impairments. This current research addresses the viability of using physically realistic 3D renderings of public spaces under artificially induced blur in place of the more resource intensive testing in the real spaces themselves while participants wear blurring goggles. In addition, this research illustrates the efficacy of a model that predicts portions of public scenes that an observer with simulated vision impairment would presumably fail to detect by comparing the predictions of missed space geometry to actual geometry detection failures by observers with simulated impairments. Lastly, this work also addresses how well simulated Low Vision observers can categorize the contents of scenes. Observer categorization rate is compared to several image metrics and the results indicate that average classification rate across Low Vision simulations can be predicted very well by knowing the averages of several different image metrics within each of the acuity blocks. Chapter 1 of this dissertation is a literature review necessary for understanding the background of the state of the art of this research and an overview of the research itself. In Chapter 2, an experiment is described in which object visibility was tested in a virtual environment with the goal of validating the use of 3D renderings as substitutive stimuli via comparing performance between the real and digital version of the same task (Bochsler et al., 2013). The objects were ramps, steps, and flat surfaces. Participants were normally sighted young adults who viewed either blurred or unblurred images. Images were blurred using a Gaussian filter on a Sloan Chart calibrated for the viewing distance of the experiment. Patterns of object identifications and confusions between the digital and physical versions of the task were highly similar. It is very likely that 3D renderings of public spaces when used in psychophysical tasks are effective substitutive stimuli for real spaces in object detection tasks. Avenues for parametric manipulations that might strengthen the argument are also explored. Chapter 3 extends the use of physics based 3D renderings to simulations of visual impairment (Thompson et al, 2017; https://github.com/visual-accessibility/deva-filter). A model of visual impairment was used to simulate 3D renderings of public spaces under increasing levels of impairment. Participants were then asked to draw the edges and contours of objects in these simulations under several separate task conditions: draw the edges of doors, stairs, obstacles, or floor-wall-ceiling connections. As simulations of visual impairment deepened, observers struggled to find the correct object contours in each of the tasks. Also, as the simulated impairments deepened, observer data often more closely matched the predictive model: a system that puts a premium on sudden changes in luminance contrast. In the absence of context and meaning, simulated Low Vision observers tend to make false positive geometrical edge identifications when a scene has non- accidental incidences of strong luminance contrast edges such as bars of light and shadows. The predictive power and utility of the model for simulating visual impairment is also discussed. Chapter 4 contains a pilot experiment which seeks to understand how well simulated Low Vision observers can classify the category of blurry scenes shown to them. Observers were asked to perform a three alternative forced choice task where they must identify if an image is one of three scenes, and observers’ classification accuracy was tracked across acuity level simulations. Several image metrics were calculated and regressed against classification accuracy of either single scenes or classification accuracy per acuity block. It was found that average classification accuracy within an acuity block could be predicted by knowing any one of several average image metrics of scenes within blocks and when regressed across acuity levels