3 32 3. Towards a task-based computational evaluation benchmark Software prototyping: edge-detection threshold In prosthetic vision research, finding the optimal image processing settings requires the evaluation of a large parameter space through experimentation. Even relatively basic image processing strategies, such as edge-detection preserve very different levels of information, depending on the configuration parameters (seeFigure3.4). Figure 3.4: Example visualization of the environment (left-most image: original frame), processed with different edge-detection thresholds (from left to right: 5, 100 and 500). As a basic case-study, this experiment explores whether the proposed RL framework can evaluate the effect of different edge-detection thresholds on the mobility performance with SPV. Making use of the simulation software from the baseline study (de Ruyter van Steveninck et al., 2022b), we created a contour-based phosphene representation with Canny edge detection (Canny, 1986). We trained three virtual implant users (RL agents with randomly initialized network parameters) per condition, testing four different values for the upper threshold parameter, namely 1, 10, 100 and 100 (the lower threshold was fixed at half of the upper threshold). Again, the number of box-collisions and the obtained reward were taken as primary measures of the performance. Hardware prototyping: number of electrodes For obtaining a higher resolution of the prosthetic vision (i.e., increasing the number of phosphenes), implantation of more electrodes is required. The implanted number of electrodes therefore has an important effect on the quality of the phosphene vision (see Figure3.5). Figure 3.5: Example of the environment observed at different phosphene resolutions. From left to right: original frame, 18×18 phosphenes, 34×34 phosphenes and 50×50 phosphenes). This experiment explores whether our RL benchmark can quantify the task-based functional quality of (simulated) prosthetic vision at different phosphene resolutions. This experiment is a virtual replication of the real-world experiments in the prior baseline
RkJQdWJsaXNoZXIy MTk4NDMw