2.4. Discussion 2 19 Figure 2.6: Results of the SharpNet trials with deep-learning based surface boundary prediction versus edge detection with the Canny algorithm. Standardized trial duration (left), standardized number of collisions (middle) and standardized subjective rating (right) are plotted for two phosphene resolutions and two levels of environmental complexity. The complex environment contained additional background and surface textures where the simple environment consisted of plain cardboard boxes. Asterisk (∗) indicates p < 0.0125, double asterisk (∗∗) indicates p < 0.0025. that trial, where one of these participants indicated specifically that forward motion and head movements sometimes contributed to the perception of depth. Some participants indicated that their strategy was dependent on the number of phosphenes for that trial. 2.4. Discussion In this study, we evaluated indoor mobility performance with a real-world simulation of prosthetic vision and contour extraction-based image processing. Besides a general evaluation of the restorable mobility performance at different phosphene resolutions, we assessed the inherent, theoretically attainable, benefits of reducing scene complexity via removal of background textures and within-object gradients. Furthermore, we investigated whether such scene simplification can be practically achieved using a deep neural network approach for surface boundary detection. In this section we provide a discussion on our findings and point out some of the current limitations and directions for future research. 2.4.1. Mobility with simulated cortical prosthetic vision The found minimal resolution of 26 × 26 phosphenes for adequate restoration mobility in a simple scene (e.g., 90.2% of obstacle avoidance with normal vision) is comparable or moderately higher than previous studies that report a minimum of 60 (Dagnelie et al., 2007), 325 (Srivastava et al., 2009)or625 (Cha et al., 1992b) simulated phosphenes. The varying results may be related to prior experience and amount of practice by the study participants, differences in the mobility task, and the implementation of the phosphene simulation. Dagnelie et al (Dagnelie et al., 2007) found that subjects with previous experience with SPV (at least 10 hours) demonstrate improved performance compared to inexperienced subjects, achieving similar results at a lower phosphene resolution. In the experiments by Srivastava et al (Srivastava et al., 2009), participants were asked for up to 9 lab-visits. In the current study, despite a majority of participants (82.4%) who indicated to have a sufficient amount of practice, we found a slight but significant improvement in average trial duration over the course of the experiment. This means that our results may be influenced by the relatively short exposure to SPV compared to aforementioned
RkJQdWJsaXNoZXIy MTk4NDMw