2 22 2. Real-world indoor mobility with simulated prosthetic vision evaluated visually and tested in behavioral pilot experiments, it might be the case that a different configuration would yield better results. The absence of improvement with the deep-learning based image processing for prosthetic vision is regrettable, given the potential that was demonstrated in other tasks such as object recognition (Han et al., 2021; Sanchez-Garcia et al., 2020) and emotion recognition (Bollen et al., 2019a; Bollen et al., 2019b). Other types of pre-processing approaches have been proposed for mobility in particular, including depth- (Barnes et al., 2011; McCarthy et al., 2015) or contour-based (Dowling et al., 2004; McCarthy et al., 2013; Vergnieux et al., 2017) rendering, where a strict surface boundary-based (i.e., ‘wireframe’) representation was found to be most effective in virtual mobility experiments (Vergnieux et al., 2017). Although the results from our physical scene simplification comparison (see previous section) support the value of this approach for lower phosphene resolutions, a feasible real-time implementation for prosthetic vision remains to be realized. 2.4.4. Limitations and future directions To our knowledge, our study is the first that directly parametrized visual complexity in a real-world mobility experiment with SPV, by adding textures on objects floors and walls. Despite these efforts in addressing real-world visual complexity of indoor environments, the current scene is still a controlled version of visual navigation. A next step towards assessment of the requirements of SPV for daily life mobility would require free navigation in interactive environments, with realistic objects and visual cues for orientation. Future work could also further investigate other more specific mobility-related problems such as stair climbing, curb following, or avoidance of elevated objects. All of these are situations that may be relevant to the clinical target population. Furthermore, note that the navigation strategies that were used by the sighted participants in this simulation study, may not capture some aspects of clinical reality. For instance, long-term cane users or people who have undergone blind rehabilitation may be very proficient in the usage of non-visual cues, such as haptic or auditory signals. It is also important to consider potential differences in perceptual learning between sighted and blind individuals, due to cortical reorganization (Horton et al., 2017). Furthermore, we did not measure eye movements and the participants in our study were unrestricted in making saccades for exploring the rendered scene. With the current state of the technology, this will not be the case for prosthetic vision users. The extent to which our results and those from other SPV studies extrapolate to visual prosthetic users remains to be tested in clinical studies. Another limiting aspect of our study concerns the realism of the phosphene simulation. Our SPV model of cortical prosthetic vision simulates a hypothetical, future visual prosthesis capable of producing idealized (small, circular, nonoverlapping) phosphenes. Our simulation did not incorporate the effects of cortical magnification (Brindley & Lewin, 1968; Srivastava et al., 2009), where size and spacing are known to increase as a function of foveal eccentricity. The same holds for dynamic effects such as phosphene fading and interactions when stimulating neighboring electrodes (Brindley & Lewin, 1968; Dobelle et al., 1974). Similar to this retinal equivalent by Beyeler et al (Beyeler et al., 2017), an interesting line of future work could focus on the development of more realistic perceptual models of cortical prosthetic vision. Lastly, some previous studies investigated the effects of scene simplification compared to direct greyscale pixel intensity mapping (e.g.,
RkJQdWJsaXNoZXIy MTk4NDMw