Information-reduction methods determine

navigation performance in simulated

prosthetic vision in virtual reality

Master Thesis

June 30, 2023

R. E. Lucas

Student Number: 6540384

Supervisor: P. C. Klink

2nd Supervisor: M. Naber

Master Artiﬁcial Intelligence

Faculty of Science

Utrecht University

Netherlands

R. E. Lucas Master Thesis

Abstract

While there are numerous innovations that help blind an partially sighted people improve

their quality of life, some tasks still remain diﬃcult. Neurotechnological innovations can re-

store a rudimentary form of artiﬁcial vision (AV) with brain implants that stimulate the brain

and create the perception of ’phosphenes’. These are dots of light with spatial locations that

correspond to the locations of the stimulating electrodes in the brain. In this paper, we inves-

tigate how diﬀerent information-reduction methods aﬀect the usability of AV for navigation by

simulating diﬀerent types of AV in a virtual reality environment. The problem with currently

existing implants is the lack of resolution. Therefore we propose a ”walkable” path implemen-

tation that guides navigating users to avoid obstacles and stay on the sidewalk. With this path,

we subtract the relevant information for navigation from the visual scene. Through an simu-

lated prosthetic vision (SPV) study we compared this walkable-path approach with two other

methods: a semantic segmentation algorithm and a combination of semantic segmentation and

the walkable-path. Diﬀerent phosphene densities where compared for each method. We found

that adding a guidance path to the visual scene, improves navigation performance. Performance

did not necessarily improve when more phosphenes where added. Subjective evaluations showed

that people preferred only having a path over having either having both a path and the environ-

ment or only the environment. These results are a step in the direction of functionally adaptable

prosthetic vision systems.

1 Introduction

Worldwide there are over 253 million blind and partially sighted people (“World Blind Union”,

n.d.). While in the past this indicated a life of poverty and few opportunities, this all changed

due to innovations and societal support for BPS people. First the invention of Braille, which made

reading possible for BPS people, and then the use of guidance animals and probing canes which

allowed them to transport on their own. Furthermore, research concerning assistive technologies

has increased over the past years (Bhowmick & Hazarika, 2017). Technological advances have

been made through the creation of Voice Over and apps like Seeing AI which allow BPS people

to interact with the world through their phone. These technologies also allow children to use

educational materials and interact more with classmates which helps their general development

(Mulloy et al., 2014).

Nonetheless, while these assistive devices increase support and quality of life (Lancioni & Singh,

2014), they do not improve vision in the literal sense. There are still many struggles that BPS people

face in their daily life. For example, smartphone use still has its challenges due to the mobile touch

screen interfaces and small keyboards (Rodrigues et al., 2020). Navigation and mobility can also

be diﬃcult because they do not have their vision to guide them and they must use other senses

to determine where obstacles are and which road they should take (Giudice, 2018; Kemp, 1981).

Besides practical problems, BPS people also have a higher risk for depression and other social

problems (Kemp, 1981; Koenes & Karshmer, 2000). The improvement of actual vision could be

beneﬁcial for BPS people to help them overcome some practical struggles and thereby improve their

mental state.

That is why there is an increasing interest in the development of Artiﬁcial Vision (AV) (Bertozzi

et al., 2002; Fernandes et al., 2012; Humayun & de Juan, 1998; Soria et al., 2006). The applications

of AV are diverse; it can be applied in the development of robots or road vehicles (Bertozzi et al.,

2002; Soria et al., 2006), but more importantly, AV can also be used in the regaining of sight for

BPS people by implementing it in neurotechnological visual prosthetic devices (Fernandes et al.,

2012).

In this paper, we will focus on the use of AV to restore a rudimentary form of sight in BPS

R. E. Lucas Master Thesis

people. We will mainly focus on the application of AV for navigation and investigate the eﬃciency

of several algorithms and AV-parameters.

1.1 Theoretical Background

The application of AV in BPS people is typically achieved through electrical stimulation. Hu-

mayun and de Juan (1998) found that the electrical stimulation of the visual cortex resulted in

the appearance of blobs of light, which are called phosphenes. These phosphenes could be used

to compensate for the lack of photoreceptors in patients (Humayun & de Juan, 1998). With the

discovery of these phosphenes, visual prostheses could be developed. Visual prostheses are devices

that can evoke a visual percept through diﬀerent stimulation methods, such as the use of electrical

stimulation (Fernandes et al., 2012) or optogenetics (Barrett et al., 2014). There are currently

roughly two diﬀerent types of visual prostheses: retinal prostheses and cortical prostheses. Retinal

prostheses evoke percepts by stimulating the retinal neurons to compensate for damaged photore-

ceptors (Pio-Lopez et al., 2021). Cortical prosthesis, on the other hand, stimulate the visual cortex

to evoke visual percept (Liu & Humayun, 2014). Humayun and de Juan (1998) discussed that when

these visual prostheses elicit meaningful visual percepts, they could help BPS people by restoring

some of their vision.

Since then, extensive research has been conducted for the development of useful visual prostheses

(Chen et al., 2009; De Ruyter Van Steveninck, G¨u¸cl¨u, et al., 2022; De Ruyter Van Steveninck, Van

Gestel, et al., 2022; Dobelle et al., 1974). In 1974, two participants were implanted with an early

visual prosthesis (Dobelle et al., 1974). Both participants were able to recognize simple patterns,

such as letters. In 2012, the recognition of letters improved further, and the prostheses could even

be used for mobility and orientation (Fernandes et al., 2012).

The improvement of mobility and orientation is an important aspect of the development of AV

and assistive technologies. This because navigation is a diﬃcult task for BPS people (Giudice, 2018;

Kemp, 1981). They often need the assistance of sighted people or a guidance dog (Bousbia-Salah

et al., 2007). There are often obstacles that need to be avoided, but that should be detected ﬁrst, a

feat that strongly relies on vision in sighted people. When using a probing cane, the user can ﬁnd

obstacles that are close by, but they cannot know what obstacles are further ahead (Bousbia-Salah

et al., 2007). However, with the use of phosphenes, the coming of obstacles further ahead could be

anticipated and avoided (De Ruyter Van Steveninck, Van Gestel, et al., 2022).

Nevertheless, the problem that often arises with this type of phosphene vision is that its ef-

fectivity depends on the number of phosphenes, which is determined by the number of implanted

electrodes. However, the space that can be stimulated with implanted electrodes in the human

primary visual cortex is very limited (van der Grinten et al., 2022). It is therefore important to

determine the desired amount of phosphenes before implanting any electrodes in the brain of a BPS

person. This because the placement of these visual prostheses can be very invasive and in the past

some implementations where experienced as rushed and ill prepared (Chen et al., 2009; Dobelle

et al., 1974; Fernandes et al., 2012). A solution for this problem is the use of simulated prosthetic

vision (SPV). Phosphene conﬁgurations can be tested noninvasively with SPV experiments that

are run with sighted participants. In such experiments, sighted people experience phosphene or

prosthetic vision through a simulation. This can for instance be done using a Virtual Reality (VR)

setup or by navigating on a computer through a virtual environment (Bollen et al., 2019; De Ruyter

Van Steveninck, Van Gestel, et al., 2022). These experiments can then be used to evaluate the

minimum requirements that are needed to restore a certain ability (Vergnieux et al., 2017). In gen-

eral, such simulations can thus help with the development of prostheses by determining the optimal

number, placement, and processing of phosphenes and electrodes without damaging a participant

R. E. Lucas Master Thesis

(Chen et al., 2009).

While SPV is a commonly used study method, it also has some drawbacks. One of the problems

is that these simulations often lack realism and biological plausibility (van der Grinten et al., 2022).

For example, the interaction between phosphenes in SPV studies is not the same as in real prostheses

(Chen et al., 2009). Furthermore, the temporal dynamics between phosphenes needs to reﬂect the

delay that can occur between the onset and oﬀset of a phosphene (Dobelle et al., 1974). Namely,

the phosphene response and duration, change depending on the stimulus presentation (Chen et

al., 2009). Due to these shortcomings, some current results of SPV studies cannot be applied to

prosthetic vision. This creates a problem for the development of visual prostheses. If we cannot

properly study a prosthesis without invasively stimulating a person’s brain, we might need to look

at other options to improve the life of BPS people.

The problems with AV and mainly SPV studies were also noted by van der Grinten et al. (2022).

As said before, SPV studies lack biological plausibility. Chen et al. (2009) noted that for SPV studies

to be relevant, it is important that simulated phosphenes are represented as those evoked by a real

visual prosthesis. That is why van der Grinten et al. (2022) developed a biologically plausible

phosphene simulator. This simulator allows phosphene vision to take several characteristics of

the cortex into account which results in phosphenes that should look very similar to those evoked

by electrical cortical stimulation. It works in real-time and therefore can be used in behavioural

experiments such as SPV studies (van der Grinten et al., 2022). Using this simulator in SPV studies

will make the results a much stronger prediction for real visual prostheses.

While this simulator helps to study prostheses, the limited number of phosphenes relative to the

rich visual scenes we typically perceive still causes a substantial proportion of information to get lost

(Sanchez-Garcia et al., 2020). This is a problem for more complex tasks, such as navigation, which

require more in-depth interpretations and exact visual cues of information (Sanchez-Garcia et al.,

2020). Simple environments can be expressed in phosphene simulations, but more complex and

real-life environments are still too complex (De Ruyter Van Steveninck, Van Gestel, et al., 2022).

That is why helpful information reduction methods need to be researched and used in (simulations

of) visual prostheses.

Diﬀerent algorithms have been used and developed to extract information from a visual scene.

A commonly used algorithm is an edge-detection algorithm. This algorithm extracts visual gra-

dients from all areas of the visual scene (De Ruyter Van Steveninck, Van Gestel, et al., 2022).

The advantage of this method is its ability to exclude noise, which is useful in real-world images

(Truchetet et al., 2001). However, this method subtracts a lot of information and requires more

phosphenes to capture the environment compared to a method where less information is subtracted.

An alternative is a surface-boundary algorithm (De Ruyter Van Steveninck, G¨u¸cl¨u, et al., 2022).

A strict version of this algorithm removes all within-surface information and background textures

from a visual scene. With this simpliﬁcation, a trade-oﬀ is created between interpretability and

informativity (De Ruyter Van Steveninck, Van Gestel, et al., 2022). Another algorithm that can be

used to visualize a scene is semantic segmentation. Sanchez-Garcia et al. (2020) clustered informa-

tion based on semantic meaning. This way objects fall under a semantic category and correspond

to a certain meaning. In their study, they combined the use of instance segmentation with the use

of semantic segmentation. This to both highlight the useful areas and show the edges of objects

and environments. Using this algorithm improved the recognition of objects and rooms compared

to a no processing method which caused an overload of information (Sanchez-Garcia et al., 2020).

De Ruyter Van Steveninck, Van Gestel, et al. (2022) did not ﬁnd any improvement in their study on

mobility performance when only using an information reduction algorithm. Sanchez-Garcia et al.

(2020), on the other hand, found that recognition performance improved when combining an infor-

mation reduction algorithm with the important aspects in the scene. This means that highlighting

R. E. Lucas Master Thesis

useful parts of the environment and subtracting only useful information can improve performance.

More studies have been conducted to test information reduction. Vergnieux et al. (2017) studied

navigation and mobility. They focussed on way ﬁnding and compared diﬀerent methods of infor-

mation display to see how well people could identify landmarks and remember the map of unknown

environments. They found that only rendering the edges of the environment was beneﬁcial for the

performance of participants. Showing that minimum information is very suﬃcient to ﬁnd one’s

way in an unknown environment. Fornos et al. (2005) studied the minimization of information in

reading. By minimising the sections shown of a text, participants were more able to read the text

(Fornos et al., 2005).

Lastly, Bollen et al. (2019) used simulated phosphene vision for emotion recognition. They com-

pared an edge detection algorithm to a more simple but still powerful image processing algorithm.

By reducing the amount of information, they found that the accuracy of emotion recognition was

saturated with a grid of 5k phosphenes. On the other hand, using the edge detection algorithm,

saturation was only found around 10k phosphenes. This illustrates that when only useful informa-

tion is used, in this case, the mouth, eyes, nose and facial contours, fewer phosphenes are necessary

to reach maximal performance.

1.2 This Paper

In this paper, we will further investigate the use of artiﬁcial vision in the improvement of

navigation for blind and partially sighted people. Since the most prominent current problem seems

to be the limited amount of information that can be represented with phosphene vision, we propose

a diﬀerent method to process visual input, namely a ”walkable path”. The walkable path is a line

that can be constructed from phosphenes and that can indicate the path where the user can walk

on the sidewalk without hitting obstacles. This makes optimal use of a low phosphene resolution,

while still enabling people to navigate through complex scenes.

We conducted an SPV experiment to test the performance on a navigation task to evaluate

this method. We will compare three information-reduction methods for processing visual input to

evaluate the diﬀerence in performance and experience: 1) a semantic segmentation (SS) method,

2) a walkable path (WP) method, and 3) a combination of these two methods (SW).

We hypothesize that the walkable path improves navigation performance, and we thus expect

a higher performance in the methods where a walkable path is present (WP & SW) compared to

the condition where this path is absent (SS). This expectation mainly stems from the fact that

limited information can increase performance (Bollen et al., 2019; De Ruyter Van Steveninck, Van

Gestel, et al., 2022). Furthermore, we expect the subjective evaluation of phosphene vision to be

more positive in the conditions where a walkable path is present because the information is more

focused on the task. We expect a subjective preference for the SW method to the WP method,

since we think that having the environment present may feel more natural and gives a reference of

movement.

We will also compare diﬀerent phosphene densities. For each method we compare the perfor-

mance across simulations with 64, 625 or 1000 phosphenes. We expect performance to improve

with higher phosphenes numbers.

2 Methods

Before starting the data collection an Ethics and Privacy Quick Scan was performed. This Scan

of the Utrecht University Research Institute of Information and Computing Sciences classiﬁed this

R. E. Lucas Master Thesis

research as low risk. No full ethics review or privacy assessment was required. This research was

executed according to the protocol of the Faculty of Social Sciences.

2.1 Participants

Data was collected from 20 participants: 10 Males and 10 females aged 19-49 years (M = 25.1,

SD = 6.37). They participated on a voluntary basis and could withdraw at any moment during the

experiment. They were, if applicable, compensated with student credit. Overall participants did

not have any visual impairments (contact lenses and glasses were allowed). One participant had

amblyopia but did not show to be an outlier. Nine participants had some experience with Virtual

Reality (VR) and twelve participants had no experience with Virtual Reality. Seven participants

described themselves as gamers. Two female participants were excluded from the study because

they could not complete the experiment due to extreme nausea in the VR environment.

2.2 Materials

2.2.1 VR

This experiment was conducted using a Virtual Reality (VR) setup. We used an HTC VIVE

PRO headset with a portable battery. The headset had a 1440x1600 resolution per eye, 98

◦

hor-

izontal Field of View and a refresh rate of 90 Hz. The headset had a wireless connection to two

base stations that were used to track the movement of the headset and controllers. The tracking

frequency was 1000 Hz. There were two controllers with multiple buttons. In this experiment we

used the track pad on top of the controllers and the index trigger on the bottom of the controllers.

2.2.2 Unity

The experiment was built in Unity

using the Unity documentation for help (Technologies,

n.d.). The base of the environment was built by Art Equilibrium (2020) and represented the streets

of New York. This environment was adapted for this experiment by writing additional scripts for

an experimental pipeline and by adding some objects in the environment. The adaptations for the

working of the simulator and diﬀerent shaders for this environment were done by De Ruyter van

Steveninck (2023).

Shaders To create the black, white edge contrast, an edge detection shader was used. This shader

was created by van der Grinten et al. (2022). It uses a Sobel ﬁlter to create the edges and does

not have a smoothing step. In ﬁgure 1b sub-panel (I) a representation of the world with only edge

detection is shown. The other ﬁgures show the world with the each information-reduction methods

and edge detection.

Semantic Segmentation The contrast between objects was created with semantic segmentation.

Each object was labelled to one of the following categories: plants, cars, houses, road, probs, signs,

fences, walkable path, or default.

Each category contained objects with corresponding semantic meanings. Every category had a

diﬀerent colour and edges were created between objects of a diﬀerent category. This is shown in

ﬁgure 1a sub-panel (II).

R. E. Lucas Master Thesis

Figure 1: Visual representation of the information-reduction methods applied to the environment

before and after applying edge detection

(a) Before applying the edge detection algorithm

I) Before any processing,

II) the SS method,

III) the SW method,

IV) the WP method

(b) After applying the edge detection algorithm

I) only edge detection,

II) the SS method,

III) the SW method,

IV) the WP method

Walkable Path Throughout the environment, a line was drawn that avoided obstacles and kept

participants on the sidewalk. When walking on this path and following it, participants should be

able to walk a perfect route.

Under the hood, this worked the same as the Semantic Segmentation. However, there were only

two categories and thus two colours. So, the edges existed between the lines of the path and the

rest of the environment. This is shown in ﬁgure 1a sub-panel (IV).

2.2.3 Phosphene Simulator

For the phosphene simulator, we used an implementation of the simulator by van der Grinten

et al. (2022). This is a biological plausible simulator that ensures that the results of this study will

be informative for real cortical implants. Three diﬀerent ﬁles were generated with either 64, 625

or 1000 phosphenes within a ﬁeld of view of 50 degrees. The locations of these phosphenes were

determined in a uniform randomized matter using polar coordinates. This can be seen in ﬁgure 2.

The size of the receptive ﬁeld of the phosphene was set to 5. This was relative to the phosphene

size which was recorded in dynamic visual acuity. Phosphenes became smaller and denser when

they were in the fovea and bigger once they spread out towards the peripheral ﬁeld. Phosphenes

became activated when edges became within the ﬁeld of view.

2.2.4 Questionnaire

A questionnaire was used to evaluate the subjective experience of the participants. We measured

how easy they found the task, how mentally and physically challenging it was, whether they felt

stressed or comfortable, if they were aware of the environment and if they needed more guidance.

Participants had to respond on a 7-point Likert scale how much they agreed with the statement.

1 indicated they strongly agreed, 4 indicated that they neither agreed nor disagreed and 7 indicated

that they strongly disagreed.

The following statements were used to evaluate the subjective experience:

R. E. Lucas Master Thesis

Figure 2: Phosphene densities in 50 degrees FoV

Note. I) 64 phosphenes, II) 625 phosphenes and III) 1000 phosphenes

1. The overall task I was assigned was easy to complete.

2. The navigation task was mentally challenging.

3. The navigation task was physically challenging.

4. I felt stressed while completing the navigation task.

5. I felt comfortable while navigating in the VR environment.

6. I was aware of my surroundings.

7. More guidance during the task would be beneﬁcial for me.

8. The choices I made throughout the task were based on my understanding about the environ-

ment.

After collecting the responses, some questions were re-scaled so that an overall score could be

interpreted. Question 2, 3, 4, 7 and 8 were re-scaled by replacing the low numbers with correspond-

ing high numbers. After that an overall score could be calculated. This score represents a positive

or negative experience, where a higher score means a more negative experience and a lower score a

more positive experience.

2.3 Conditions & Task

All participants experienced all three information-reduction methods: semantic segmentation

(SS), walkable path (WP) and these two methods combined (SW). In the SS condition the partici-

pant had to ﬁnd their own way and deduce from the environment where they could and could not

walk. In the WP condition participants only had a path to guide them. This path showed them

where they could walk to stay on the sidewalk or crossroad and without hitting obstacles. In the

SS condition participants had the path to guide them but could also interpret the environment in

their own way.

For each method three diﬀerent phosphene concentrations where compared: 64, 625 or 1000

phosphenes. So, in total there were 9 diﬀerent conditions that all participants experienced. Partic-

ipants were divided into six groups. Each group had a diﬀerent order of the information-reduction

methods. So, they would either start with SS, WP, or SW. This was done so that the experiment

was counterbalanced to avoid recency and learning eﬀects. An overview of the diﬀerent conditions

can be found in ﬁgure 3.

R. E. Lucas Master Thesis

Figure 3: Diﬀerent phopshene densities for each information-reduction method

Note. The ﬁgures (A, B, C) on the ﬁrst row contain 64 phosphenes. The ﬁgures (D, E, F) on the

second row contain 625 phosphenes. The ﬁgures (G, H, I) on the third row contain 1000 phosphenes.

The ﬁrst column (A, D, G) presents the WP method, the second column (B, E, H) the SS method

and the third column (C, F, I) the SW Method. For panels F and I the path is accentuated with

a line.

In each trial, participants were asked to complete a navigation task. They either had to get

some bread at the bakery, post a letter, ﬁnd a taxi, or throw away the trash as can be seen in ﬁgure

4. Each task had a diﬀerent starting point at which participants heard which goal they had to ﬁnd.

An audio voice guided them in the right direction. This voice told them to either go left or right,

or whether the destination was on their left or right side. If there were no directions, participants

were instructed to walk straightforward. When arriving at the ﬁnish, participants had to either

bump or walk into the ﬁnish.

2.4 Procedure

After entering the lab, the participants were asked to sign an informed consent and got a brief

instruction of the task. Then the VR headset was put on their head and adjusted to the right

settings.

First participants had to type in their participant number and group number in the menu.

After that the participant was allowed to practice in the VR environment with guidance of the

R. E. Lucas Master Thesis

Figure 4: End goals of the four diﬀerent trials

Note. I) The Bakery, II) The Postbox, III) The Taxi, IV) The Trash

experiment leader. This was done to make sure that they felt comfortable and also to show the

diﬀerent settings they would experience.

When the participant felt comfortable enough in the environment, they started practicing the

routes of the four diﬀerent trials. After practicing these routes, the real experiment began. For

each condition, the participant completed four trials. Each trial took a maximum of 120 seconds.

After these four trials, the participant was asked a few questions about their experience with this

navigation method. During the experiment, the experiment leader monitored if participants felt

nauseous and if they needed a break. The experiment leader also helpt redirecting participants if

they got lost in the environment. In total, participants completed 36 trials and 9 questionnaires.

After completing the experiment, the VR headset was removed and participants where, when

applicable, compensated for their time.

2.5 Data Collection & Analysis

Data was logged for each trial. We logged the time it took participants to complete a trial,

the number of times they bumped into obstacles and the number of times they deviated from the

sidewalk.

We then ﬁrst tested the impact of the type of information-reduction method and the number of

phosphenes on the combination of the three dependent variables with a MANOVA Wilks test. After

that, we tested the impact of the methods and number of phosphenes on the individual dependent

variables with multiple ANOVA tests. We also compared the diﬀerence between diﬀerent number of

phosphenes conditions within a method. Subjective responses were also compared across diﬀerent

methods and number of phosphenes.

Whenever we found a signiﬁcant eﬀect in these tests, we performed post-hoc tests with a Tukey

correction to study the diﬀerence across groups.

All statistical tests were done using JASP (JASP Team, 2023).

R. E. Lucas Master Thesis

3 Results

The tables of these tests can be found in Appendix A. We also performed Bayesian ANOVAs

to test the robustness of these results. Those ﬁndings can be found in Appendix B.

3.1 Descriptive Statistics

The descriptive statistics for each number of phosphenes can be found in table 1. The descriptive

statistics for each information-reduction method can be found in table 2.

Table 1: Descriptive Statistics for the diﬀerent numbers of phosphenes

trialTime obstacleBumpCount roadBumpCount

64 625 1000 64 625 1000 64 625 1000

Valid 212 216 216 212 216 216 212 216 216

Missing 4 0 0 4 0 0 4 0 0

Mean 100.258 78.523 71.133 34.736 24.611 21.236 4.302 3.449 2.861

Std. Deviation 26.280 31.958 29.129 21.951 17.370 15.793 4.809 5.153 4.213

Minimum 35.681 27.306 21.944 4.000 2.000 2.000 0.000 0.000 0.000

Maximum 122.517 122.518 122.510 101.000 89.000 81.000 30.000 43.000 22.000

Table 2: Descriptive Statistics for the diﬀerent information-reduction methods

trialTime obstacleBumpCount roadBumpCount

SS SW WP SS SW WP SS SW WP

Valid 212 216 216 212 216 216 212 216 216

Missing 4 0 0 4 0 0 4 0 0

Mean 92.131 81.998 75.635 32.703 25.611 22.231 5.877 3.181 1.583

Std. Deviation 28.368 32.071 32.355 20.533 19.750 16.169 5.408 4.512 3.100

Minimum 31.785 21.944 23.666 4.000 2.000 3.000 0.000 0.000 0.000

Maximum 122.514 122.518 122.514 101.000 89.000 82.000 43.000 27.000 22.000

3.2 General Eﬀect

First we looked at the general eﬀect on trial time, number of obstacle bumps and number

of road bumps. We found that phosphene number (Wilks’ Λ(2, 608) = 0.820, p < .001) and

information-reduction method (Wilks’ Λ(2, 608) = 0.827, p < .001) both had a signiﬁcant eﬀect on

the dependent variables combined. The type of trial also had a signiﬁcant eﬀect on the dependent

variables combined (Wilks’ Λ(3, 608) = 0.831, p < .001). We are not interested in the eﬀect of trial

type on performance but take it into our analysis to explain some of the variance in our data. Fur-

thermore, the interaction between information-reduction method and phosphene numbers (Wilks’

Λ(4, 608) = 0.963, p = 0.028) and the interaction between trial type and information-reduction

method (Wilks’ Λ(6, 608) = 0.937, p = 0.002) showed to be signiﬁcant. The diﬀerent interactions

between phosphene number and information-reduction method did not have a signiﬁcant eﬀect on

the number of obstacle bumps or the time needed to complete a trial.

R. E. Lucas Master Thesis

After looking at the general eﬀect on performance, we zoomed in on the individual impact of

the independent variables on the dependent variables. All eﬀects are visualized in ﬁgure 5, 6, 7.

Figure 5: Average number of obstacle bumps

Note. A higher score corresponds to a more obstacle bumps.

Figure 6: Average number of obstacle bumps

Note. A higher score corresponds to a more road bumps.

Figure 7: Average time needed to complete a trial

Note. A higher score corresponds to a more trial time.

3.3 Phosphene Number

The number of phosphenes had a signiﬁcant eﬀect on the number of obstacle bumps (F (2, 632) =

36.539, p < 0.001), the number of road bumps (F (2, 632) = 6.382, p = 0.002) and the time it took

to complete a trial (F (2, 632) = 66.094, p < .001).

R. E. Lucas Master Thesis

We performed a post hoc test to test the diﬀerences between phosphene densities on obstacle

bumps. We found that 64 phosphenes diﬀered signiﬁcantly from 625 phosphenes (t = 5.886,

p < .001) and 1000 phosphenes (t = 7.825, p < .001). When using 64 phosphenes participants

had on average 10.246 more obstacle bumps compared to using 625 phosphenes and 13.621 more

obstacle bumps compared to using 1000 phosphenes. 625 phosphenes and 1000 phosphenes did not

diﬀer signiﬁcantly (t = 1.948, p = 0.126).

We also performed a post hoc test to see the diﬀerence between phosphene densities on number

of road bumps. We found that 64 phosphenes did not diﬀer signiﬁcantly from 625 phosphenes

(t = 2.116, p = 0.087). 64 phosphenes did diﬀer signiﬁcantly from 1000 phosphenes (t = 3.501,

p = 0.001). Compared to 1000 phosphenes participants had on average 1.486 more road bumps when

using 64 phosphenes. 625 phosphenes and 1000 phosphenes did not diﬀer signiﬁcantly (t = 1.392,

p = 0.346).

Lastly, we performed a post hoc test to see the diﬀerence between phosphene densities on time

needed to complete a trial. We found that 64 phosphenes diﬀered signiﬁcantly from 625 phosphenes

(t = 7.953, p < .001) and 1000 phosphenes (t = 10.646, p < .001). 625 phosphenes and 1000

phosphenes also diﬀered signiﬁcantly (t = 2.706, p = 0.019). Using 1000 phosphenes participants

were on average 7.390 seconds faster compared to using 625 phosphenes and 29.215 seconds faster

compared to using 64 phosphenes. Using 625 phosphenes, participants were on average 21.825

seconds faster compared to using 64 phosphenes.

3.4 Information-reduction methods

The type of information-reduction method had a signiﬁcant eﬀect on the number of obstacle

bumps (F (2, 632) = 21.433, p < .001), the number of road bumps (F (2, 632) = 54.095, p < .001)

and the time it took to complete a trial (F (2, 632) = 20.466, p < .001).

We performed a post hoc test to see the diﬀerence between methods on number of obstacle

bumps. We found that SS method diﬀered signiﬁcantly from the SW method (t = 4.165, p < .001)

and the WP methods (t = 6.107, p < .001). With using the SS method participants had on

average 7.251 more obstacle bumps compared to using the SW method and 10.631 more obstacle

bumps compared to using the WP method. The SW method and the WP method did not diﬀer

signiﬁcantly (t = 1.951, p = 0.125).

We again performed a post hoc test to see the diﬀerence between methods on the number of

road bumps. We also found that SS method diﬀered signiﬁcantly from the SW method (t = 6.392,

p < .001) and the WP methods (t = 10.155, p < .001). The SW method and the WP method also

diﬀered signiﬁcantly (t = 3.782, p = 0.123). Using the WP method, participants had on average

1.597 fewer road bumps compared to using the SW method and 4.310 fewer road bumps compared

to using the SS method. Using the SW method, participants had on average 2.713 fewer road

bumps compared to using the SS method.

The last post hoc test was performed to see the diﬀerence between methods on the time needed

to complete a trial. We also found that SS method diﬀered signiﬁcantly from the SW method

(t = 3.780, p < .001) and the WP methods (t = 6.099, p < .001). Participants where 10.374

seconds slower when using the SS method compared to the SW method and 16.737 seconds slower

when using the SS method compared to the WP method. The SW method and the WP method

did not diﬀer signiﬁcantly (t = 2.330, p = 0.052).

Lastly, we zoomed in on the diﬀerent information-reduction methods to look at the diﬀerence

between number of phosphenes within a method.

R. E. Lucas Master Thesis

3.4.1 Semantic Segmentation

For the SS method we found that there was a signiﬁcant eﬀect of phosphene number on the

number of obstacle bumps (F (2, 209) = 10.846, p < 0.001) and the trial time (F (2, 209) = 15.242,

p < .001). There was no signiﬁcant eﬀect on the number of road bumps (F (2, 209) = 1.670,

p = 0.191). A post hoc test showed that there was a signiﬁcant diﬀerence between 64 phosphenes

and 625 phosphenes for both obstacle bumps (t = 3.042, p = 0.007) and time (t = 2.982, p = 0.009).

Using 64 phosphenes, participants had 10.100 more obstacle bumps and were 13.427 seconds slower

than when using 625 phosphenes. There also was a signiﬁcant eﬀect between 64 phosphenes and

1000 phosphenes both obstacle bumps (t = 4.589, p < 0.001) and time (t = 5.519, p < .001).

Using 64 phosphenes, participants had 15.239 more obstacle bumps and were 24.849 seconds slower

than when using 1000 phosphenes. There was no diﬀerence between 625 phosphenes and 1000

phosphenes for obstacle bumps (t = 1.570, p = 0.261). For time there was a signiﬁcant diﬀerence

(t = 2.574, p = 0.029). Using 1000 phosphenes, participants were 11.422 seconds faster compared

to using 625 phosphenes.

3.4.2 Walkable Path

For the WP method we found that there was a signiﬁcant eﬀect of phosphene number on the

number of obstacle bumps (F (2, 213) = 6.174, p = 0.002) and the trial time (F (2, 213) = 21.308,

p < .001). There was no signiﬁcant eﬀect on the number of road bumps (F (2, 213) = 0.679,

p = 0.508). A post hoc test showed that there was a signiﬁcant diﬀerence between 64 phosphenes

and 625 phosphenes for both obstacle bumps (t = 2.992, p = 0.009) and time (t = 5.595, p <

.001). Participants bumped on average into 7.875 more obstacles and where 27.670 seconds slower

when using 64 phosphenes compared to using 625 phosphenes. There also was a signiﬁcant eﬀect

between 64 phosphenes and 1000 phosphenes both obstacle bumps (t = 3.092, p = 0.007) and

time (t = 5.710, p < .001). Here participants bumped on average into 8.139 obstacles and where

28.241 seconds slower when using 64 phosphenes compared to using 625 phosphenes. There was no

diﬀerence between 625 phosphenes and 1000 phosphenes both obstacle bumps (t = 0.100, p = 0.994)

and time (t = 0.115, p = 0.993).

3.4.3 Semantic Segmentation with Walkable Path

For the SW method we found that there was a signiﬁcant eﬀect of phosphene number on the

number of obstacle bumps (F (2, 213) = 17.410, p < .001), number of road bumps (F (2, 213) =

7.071, p = 0.001) and the trial time (F (2, 213) = 27.515, p < .001). A post hoc test showed that

there was a signiﬁcant diﬀerence between 64 phosphenes and 625 phosphenes for both obstacle

bumps (t = 4.163, p < .001) and time (t = 5.092, p < .001). Using 64 phosphenes, participants

on average bumped into 12.764 obstacles and where 24.378 seconds slower compared to using 625

phosphenes. There was no signiﬁcant diﬀerence between 64 phosphenes and 625 phosphenes for

road bumps (t = 1.765, p = 0.184). There also was a signiﬁcant eﬀect between 64 phosphenes

and 1000 phosphenes obstacle bumps (t = 5.703, p < .001), road bumps (t = 3.758, p < .001)

and time (t = 7.218, p < .001). Using 64 phosphenes, participants on average bumped into 17.465

more obstacles, had 2.750 more road bumps and were 34.554 seconds slower compared to using

1000 phosphenes. There was no diﬀerence between 625 phosphenes and 1000 phosphenes obstacle

bumps (t = 1.540, p = 0.274), road bumps (t = 1.993, p = 0.116) and time (t = 2.126, p = 0.087).

R. E. Lucas Master Thesis

Figure 8: Average response on the questionnaire

Note. A higher score corresponds to a more negative experience, a lower score to a more positive

experience.

3.5 Questionnaire

After rescaling the data, we found that the type of information-reduction method (F (2, 1279) =

24.175, p < .001) and number of phosphenes (F (2, 1279) = 40.317, p < .001) on the response on

the questionnaire. A visualization of the data can be found in ﬁgure 8. A post hoc test showed that

there was a signiﬁcant diﬀerence between 64 phosphenes compared to 625 phosphenes (t = 6.724,

p < .001) and 1000 phosphenes (t = 8.537, p < . 001). On average when using 64 phosphenes,

participants were 0.824 points more negative compared to using 625 phosphenes and 1.047 points

more negative compared to using 1000 phosphenes. There was no signiﬁcant diﬀerence between

625 phosphenes and 1000 phosphenes (t = 1.824, p = 0.162). This test also showed that there was

a signiﬁcant eﬀect between the SS method compared to the WP method (t = 6.515, p < .001) and

the SW method (t = 5.303, p < .001). On average, using the SS method, participants were 0.650

points more negative compared to using the SW method and 0.789 points more negative compared

to using the WP method. There was no signiﬁcant diﬀerence between the WP method and the SW

method (t = 1.130, p = 0.496).

4 Discussion

In this study we explored the eﬃcacy of diﬀerent forms of artiﬁcial vision for navigation. More

speciﬁcally, we studied the impact of diﬀerent information-reduction methods to process visual

input and diﬀerent phosphene densities on the performance of participants on a navigation task.

We compared three methods namely, semantic segmentation (SS), a walkable path (WP) and the

combination of these two (SW). For each method, we compared three diﬀerent phosphene densities:

64, 625 and 1000 phosphenes. We evaluated each condition on three variables: number of obstacle

bumps (obstacle bumps), number of road bumps (road bumps) and the time needed to complete a

trial (trial time). We also asked participants to rate their qualitative experience of these conditions

and to state their personal preference for a method. This simulated phosphene vision experiment

thus tested how eﬃciently people navigate in these conditions and how they experienced these

diﬀerent visual experiences.

The results indicate that the overall performance is aﬀected by the number of phosphenes and by

R. E. Lucas Master Thesis

the type of information-reduction method. This means that the mean performance on each method

and each phosphene density was signiﬁcantly diﬀerent from the other methods and densities. We

looked at each individual performance variable to see how each method and density diﬀered from

the others.

First, we zoomed in on the impact of the number of phosphenes on the individual perfor-

mance variables. We found that using 64 phosphenes leads to more road bumps, obstacle bumps

and a higher trial time compared to using 625 phosphenes or 1000 phosphenes. When using

625 phosphenes, participants needed more time to complete a trial then when they used 1000

phosphenes. Using either 625 or 1000 phosphenes showed no diﬀerence in the number of obsta-

cle or road bumps. This can be explained by the fact that a higher resolution can lead to more

detailed visual percepts. With only 64 phosphenes a user has a harder time to determine what

objects are in front of them or when they are walking onto the road. This because a user cannot

see everything around them when using just a few phosphenes. With 625 or 1000 phosphenes this

becomes increasingly easier. Objects can be recognized easier and with 1000 phosphenes, users can

discriminate better between the road and sidewalk because they have more phosphenes to look at

both obstacles and the road. Furthermore, when using more phosphenes a user can become more

conﬁdent when walking, which will lead to a faster trial time.

Secondly, we looked at the impact of the diﬀerent information-reduction methods on the indi-

vidual performance variables. We found that the SS method leads to more obstacle bumps, road

bumps and completion time compared to the SW and WP method. Using the WP method, partic-

ipants bumped less into the road and needed less time to complete a trial. There was no diﬀerence

in the number of obstacle bumps when using the WP or SW method. When using the SS method,

users have to determine their own path while visual input is limited. Users have to decide what the

diﬀerence is between the road, an obstacle and the sidewalk but also decide if they turned enough

towards the right direction. This decision making takes longer then when they have a guided path.

This decision making process is also a problem for the SW method. Users have to determine the

diﬀerence between the path and the environment. This can cause some confusion which leads to

slower times and more road bumps.

Thirdly, we dived deeper into these individual information-reduction methods to see the diﬀer-

ence in impact on performance between phosphene densities. For each method, participants needed

more time to complete a trial and bumped into more obstacles when using only 64 phosphenes.

When using the SS method, participants where the fastest when navigating with 1000 phosphenes.

There was no signiﬁcant diﬀerence in the number of obstacle bumps when using SS with 625 or

1000 phosphenes. For the SS method and WP method there was no diﬀerence in the number of

road bumps between phosphene densities. When using the SW method, participants also performed

worse on the number of road bumps when using only 64 phosphenes. More phosphenes lead to a

better performance, but there was no signiﬁcant diﬀerence on all variables between 625 and 1000

phosphenes for the SW method. This was also the case when using the WP method for the number

of obstacle bumps or trial time. As said before, with fewer phosphenes users need more time to

discriminate between obstacles and the road but also see less of the world around them which

causes a worse performance. For the SS method, we see that more phosphenes leads to a faster

performance. This can be caused by the fact that users feel more conﬁdent and thereby walk faster.

When using SW with 64 phosphenes, users had a harder time discriminating between the road and

sidewalk then when they had more phosphenes. This might have been caused by mistaking the line

that divides the road and the sidewalk for the walkable path. This causes the users to diﬀer more

from the road instead of staying on the sidewalk. When the number of phosphenes increases, this

division became more clear and thereby reduced the number of road bumps. We also see that for

most variables there was no diﬀerence between 625 and 1000 phosphenes.

R. E. Lucas Master Thesis

Lastly, we zoomed in on the average responses per question to look for an explanation for

the subjective preference of the walking path. We saw that participants found the task slightly

easier when only using the walkable path compared to also having the environment. They also felt

slightly less mentally and physically challenged and were a bit more comfortable when using only

the walkable path. This may have lead to the observed subjective preference for the walkable path

over the combination of the path and environment.

Altogether, this study shows that from those tested, the worst information-reduction method

to use for navigation is the SS method with only 64 phosphenes. The best method to use is the WP

method because there were less road bumps and less time was needed to complete a trial. For this

method, looking at the objective measurements and the questionnaire it does not matter whether

the user uses 625 phosphenes or 1000 phosphenes. However, there was an observed subjective

preference for more phosphenes over less.

In line with our expectations, we see that using a walkable path leads to better performance

compared to using no path for navigation. We see that the performance on navigation using the WP

or SW method diﬀers a bit from each other. Counter to our hypothesis, participants preferred using

only the path without the environment. This did not show from the questionnaire results in which

we only saw a more positive experience for a method with a path compared to a method without

a path. When, however, we directly asked participants which method they would prefer to use

during navigation, most preferred using only the walkable path method. We see that performance

indeed improves when adding more phosphenes but we also see that more phosphenes in this case

does not necessarily lead to a better performance.

So, the choice for the optimal information-reduction method is based on both objective and

subjective measurements. A method should be objectively one of the best performing methods.

Furthermore, it should be relatively positively experienced. When multiple methods meet these

criterion, a personal preference of the user can determine which method will be implemented in the

end.

In conclusion, the results of this study imply that using a walkable path can improve performance

on a navigation task but that performance does not keep increasing with increasing phosphene

counts. While current studies still work on the improvement of electrode placement (Abouelseoud

et al., 2022), it is useful to keep in mind that improvement of performance can saturate.

To use a walkable path for real-life navigation some extra steps are necessary in the processing

of the environment. An algorithm is needed to scan the environment and classify all obstacles

and detect the diﬀerence between the road and sidewalk. Existing algorithms can be used such

as the edge detection of semantic segmentation algorithms. A new algorithm must be designed to

determine, within this classiﬁcation, which of the spaces are unoccupied. This algorithm should

then determine whether these spaces are big enough to be walked upon by a human. Then, the

algorithm needs to draw a line through these spaces to guide the user in the environment. This

algorithm should work in real-time so that it can work in any environment.

Additionally, these results suggest that we may want to rethink how visual prosthesis are de-

signed. Traditionally, such prostheses are regarded as a unitary substitute of general vision, apply-

ing a single image processing algorithm to its input to perform the required information reduction

and generate informative, yet relatively spare, stimulated patterns. However, diﬀerent visual func-

tions may require diﬀerent information reduction strategies and we may want to emulate that in

an adaptive prosthesis. We see that using a walkable path is more beneﬁcial for the performance

on a navigation task than seeing the environment. This implies that if visual prostheses used this

walkable path, users would improve their performance with navigation, but this does not mean

that the same algorithm would be optimal for object recognition or reading.

A limitation of our study is the use of SPV. Despite using a realistic phosphene simulator

R. E. Lucas Master Thesis

(van der Grinten et al., 2022), participants still had to move through the environment by using the

controllers, because of space limitations. Participants did not actually bump into objects, so they

did not experience the somatosensory feedback and disadvantage of bumping into things. Also,

because the environment was artiﬁcial, participants could not make adequate use of their sense of

space. Having these abilities can help improve performance and make navigation easier (Giudice,

2018). This makes the navigation less realistic.

For this study, we only looked at the number of phosphenes as a parameter. However, there

are many more parameters that could be ﬁne-tuned and tested for this navigation task. The visual

ﬁeld in which the phosphenes are generated could be an interesting parameter to investigate. For

example, how would a smaller or larger visual ﬁeld impact the use of a walkable path. Furthermore,

it might be informative to look at the coverage of the visual ﬁeld. In this study we used the full

coverage the visual ﬁeld, however due to surgical limitations coverage might be restricted. Future

research should study these other parameters and ﬁne-tune them for prostheses.

Another limitation was that we did not test the eﬀects of the walkable path on depth perception.

In our VR environment participants did not have the need to use their depth perception due to

the fact that there were no staircases or height diﬀerences to take into account during navigation.

However, for real prostheses it is important to look into the impact of a walkable path on depth

perception because depth perception can be crucial aspect of sight (Zhao et al., 2018). Future

research should study the use of a walkable path with staircases and height diﬀerences to test its

impact on depth perception.

Due to some of these limitations it might be useful to also look at some alternatives for prosthe-

ses. One of the alternatives is to abandon the use of vision altogether. When vision is not available,

other senses are used to navigate through the world (Giudice, 2018). Mobility and orientation skills

can be learned to move around objects and orient oneself in an environment (Giudice, 2018). Also,

hearing can be used to compensate for the loss of vision. For example, hearing the sound of cars

when approaching a street, or the clicking sound when waiting at a traﬃc light. So instead of seeing

the environment BPS people can hear it or develop a sense of presence in the environment. An-

other solution is to use diﬀerent technologies for navigation. Tan et al. (2021) created a new type of

guidance dog, namely a drone. They used this drone to scan the environment and help BPS people

navigate through the city. The downside of these drones was that they are not weather-resistant.

The drones are very light which makes it hard for them to ﬂy in a strong wind (Tan et al., 2021).

The solution to this problem was to use stronger drones for navigation. Gonz´alez-Mora et al. (2006)

researched portable electronic devices that allow BPS people to receive spatial information from

their environment. This meant that users could ”hear” their environment and know when they

where nearby obstacles. (Bousbia-Salah et al., 2007) proposed an improved cane that could warn

users of obstacles through vibrations. However, while it senses obstacles, it cannot give the user

a sense of the environment (Bousbia-Salah et al., 2007). The problem with these alternatives is

that they can be very costly without actually restoring vision. So, using alternatives also has its

problems. That is why it remains important to further investigate the applications of AV in BPS

people.

Looking back at one of those applications, the WP can be beneﬁcial for navigation. However, as

said before, the WP method might not be optimal for other tasks. When only using this method,

users would have problems with recognizing objects, rooms or emotions. That is why we would

propose to add an adaptive switching system to a visual prosthesis. This system can switch between

diﬀerent modes depending on the setting and task that is at hand, for example, a navigation mode

for ﬁnding a friends house, a reading mode for a medicine package and diﬀerent recognition modes

for seeing the table or person in front of them. For each mode, a diﬀerent information-reduction

algorithm can be used. For example, for emotion recognition Bollen et al. (2019) method should

R. E. Lucas Master Thesis

be used. For recognizing objects and rooms, an edge detection algorithm or semantic segmentation

method could be used (De Ruyter Van Steveninck, Van Gestel, et al., 2022). This switching system

changes the focus of current research to more task speciﬁc work, looking for the optimal settings

for each task.

Before implementing the proposed switching system onto real prostheses, future research should

be conducted about its applications. For cochlear implants, switching between a quiet and noise

setting improved speech intelligibility (Hey et al., 2016). An experimental study should be con-

ducted to test whether participants are able to operate such a switching system for visual implants.

In this experiment participants could be presented with diﬀerent environments and tasks. They

should determine what setting they are present and then switch to the corresponding mode on

their system. This experiment could measure performance through completion time and correct

switches. After testing the switching system objectively, a thorough subjective evaluation should

be conducted. First an SPV study could investigate whether people enjoy using diﬀerent modes or

if they ﬁnd them tiring. Then, for example, a focusgroup could be created to evaluate how BPS

people feel about using such a system. For cochlear implants these personalized systems are already

in development. Multiple apps have been created by companies so that users can ﬁne tune and

personalize their own implants. Examples are the Nucleus

Smart app and the Baha

Smart app.

However, no research could be found that evaluated how the user experiences switching between

diﬀerent settings so an evaluating study could be beneﬁcial. This evaluation should ﬁrst happen in

an SPV study but then also for real prostheses users.

Another important aspect that should be studied, are the technical details of this switching

system. Using multiple information-reduction algorithms, will increase the processor that is needed

in an visual prosthesis. Optimization might be necessary to keep the processor compact and

portable, despite using multiple algorithms. A new study could construct this system and optimize

performance.

Lastly, another branch that could be explored within this, is to let an artiﬁcial intelligence (AI)

decide which mode is most applicable for a certain task or environment. This could be beneﬁcial if

these proposed studies show that users are unable to detect what setting they are present or what

kind of task they need to complete. An AI could detect the setting and switch to the corresponding

mode to improve user experience. This implementation of AI has already started to be research

for cochlear implants (Hey et al., 2016) but more research should be done before this could be

implemented.

In general, big steps are being made to improve vision for BPS people. We still have a long way

to go, but we are starting to ﬁnd the right path.

5 Acknowledgements

We would like to thank Jaap for providing the code of the simulator and helping with the

implementation of the simulator in Unity. Furthermore, we thank Django for providing tips for

setting up the VR setup. Lastly, thanks to Chris for supervising this thesis project.

R. E. Lucas Master Thesis

References

Abouelseoud, G., Abouelseoud, Y., Shoukry, A. A., Ismail, N. E., & Mekky, J. (2022). A Mixed

Integer Linear Programming Framework for Improving Cortical Vision Prosthesis Designs.

Social Science Research Network. https://doi.org/10.2139/ssrn.4010997

Art Equilibrium. (2020). Street New York — 3D Urban — Unity Asset Store. https://assetstore.

unity.com/packages/3d/environments/urban/street-new-york-183319#publisher

Barrett, J. W., Berlinguer-Palmini, R., & Degenaar, P. (2014). Optogenetic approaches to retinal

prosthesis. Visual Neuroscience, 31 (4-5), 345–354. https://doi.org/10.1017/s0952523814000212

Bertozzi, M., Broggi, A., Cellario, M., Fascioli, A., Lombardi, P., & Porta, M. (2002). Artiﬁcial

vision in road vehicles. Proceedings of the IEEE, 90 (7), 1258–1271. https://doi.org/10.

1109/jproc.2002.801444

Bhowmick, A., & Hazarika, S. M. (2017). An insight into assistive technology for the visually

impaired and blind people: State-of-the-art and future trends. Journal on Multimodal User

Interfaces, 11 (2), 149–172. https://doi.org/10.1007/s12193-016-0235-6

Bollen, C. J., Van Wezel, R. J. A., Van Gerven, M. A. J., & G¨u¸cl¨ut¨urk, Y. (2019). Emotion Recog-

nition with Simulated Phosphene Vision. Proceedings of the 2nd Workshop on Multimedia

for Accessible Human Computer Interfaces, 1–8. https://doi.org/10.1145/3347319.3356836

Bousbia-Salah, M., Redjati, A., Fezari, M., & Bettayeb, M. (2007). An ultrasonic navigation system

for blind people. 2007 IEEE International Conference on Signal Processing and Communi-

cations, 1003–1006. https://doi.org/10.1109/icspc.2007.4728491

Chen, S. C., Suaning, G. J., Morley, J. W., & Lovell, N. H. (2009). Simulating prosthetic vision: I.

visual models of phosphenes. Vision research, 49 (12), 1493–1506. https://doi.org/10.1016/

j.visres.2009.02.003

De Ruyter Van Steveninck, J., G¨u¸cl¨u, U., Van Wezel, R. J. A., & Van Gerven, M. A. J. (2022).

End-to-end optimization of prosthetic vision. Journal of Vision, 22 (2), 20. https://doi.org/

10.1167/jov.22.2.20

De Ruyter Van Steveninck, J., Van Gestel, T., Koenders, P., Van Der Ham, G., Vereecken, F.,

G¨u¸cl¨u, U., Van Gerven, M. A. J., G¨u¸cl¨ut¨urk, Y., & Van Wezel, R. J. A. (2022). Real-world

indoor mobility with simulated prosthetic vision: The beneﬁts and feasibility of contour-

based scene simpliﬁcation at diﬀerent phosphene resolutions. Journal of Vision, 22 (2), 1.

https://doi.org/10.1167/jov.22.2.1

Dobelle, W. H., Mladejovsky, M. G., & Girvin, J. (1974). Artiﬁcial vision for the blind: Electrical

stimulation of visual cortex oﬀers hope for a functional prosthesis. Science, 183 (4123), 440–

444. https://doi.org/10.1126/science.183.4123.440

Fernandes, R. A. B., Diniz, B., Ribeiro, R., & Humayun, M. (2012). Artiﬁcial vision through

neuronal stimulation. Neuroscience letters, 519 (2), 122–128. https://doi.org/10.1016/j.

neulet.2012.01.063

Fornos, A. P., Sommerhalder, J., Rappaz, B., Safran, A. B., & Pelizzone, M. (2005). Simulation of

Artiﬁcial Vision, III: Do the Spatial or Temporal Characteristics of Stimulus Pixelization

Really Matter? Investigative Ophthalmology Visual Science, 46 (10), 3906. https://doi.org/

10.1167/iovs.04-1173

Giudice, N. A. (2018). Navigating without vision: Principles of blind spatial cognition. In Handbook

of behavioral and cognitive geography. Edward Elgar Publishing.

Gonz´alez-Mora, J. L., Rodriguez-Hernandez, A., Burunat, E., Martin, F., & Castellano, M. A.

(2006). Seeing the world by hearing: Virtual acoustic space (vas) a new space perception

system for blind people. 2006 2nd International Conference on Information & Communi-

cation Technologies, 1, 837–842. https://doi.org/10.1109/ictta.2006.1684482

R. E. Lucas Master Thesis

Hey, M., Hocke, T., Mauger, S. J., & M¨uller-Deile, J. (2016). A clinical assessment of cochlear

implant recipient performance: implications for individualized map settings in speciﬁc en-

vironments. European Archives of Oto-rhino-laryngology, 273 (11), 4011–4020. https://doi.

org/10.1007/s00405-016-4130-2

Humayun, M. S., & de Juan, E. (1998). Artiﬁcial vision. Eye, 12 (3), 605–607.

JASP Team. (2023). JASP (Version 0.17.2)[Computer software]. https://jasp-stats.org/

Kemp, N. (1981). Social psychological aspects of blindness: A review. Current Psychological Re-

views, 1 (1), 69–89. https://doi.org/10.1007/bf02979255

Koenes, S. G., & Karshmer, J. F. (2000). Depression: A comparison study between blind and sighted

adolescents. Issues in Mental Health Nursing, 21 (3), 269–279. https://doi.org/10.1080/

016128400248086

Lancioni, G. E., & Singh, N. N. (2014). Assistive Technologies for Improving Quality of Life. https:

//doi.org/10.1007/978-1-4899-8029-8 1

Liu, W., & Humayun, M. S. (2014). Retinal prosthesis. Annual Review of Biomedical Engineering,

16, 165–187.

Mulloy, A. M., Gevarter, C., Hopkins, M., Sutherland, K. S., & Ramdoss, S. T. (2014). Assistive

technology for students with visual impairments and blindness. In Assistive technologies for

people with diverse abilities (pp. 113–156). Springer Nature. https://doi.org/10.1007/978-

1-4899-8029-8 5

Pio-Lopez, L., Poulkouras, R., & Depannemaecker, D. (2021). Visual cortical prosthesis: An elec-

trical perspective. Journal of Medical Engineering Technology, 45 (5), 394–407.

Rodrigues, A., Nicolau, H., Montague, K., Guerreiro, J., & Guerreiro, T. (2020). Open challenges

of blind people using smartphones. International Journal of Human–Computer Interaction,

36 (17), 1605–1622. https://doi.org/10.1080/10447318.2020.1768672

Sanchez-Garcia, M., Martinez-Cantin, R., & Guerrero, J. M. (2020). Semantic and structural image

segmentation for prosthetic vision. PLOS ONE, 15 (1), e0227677. https://doi.org/10.1371/

journal.pone.0227677

Soria, C. M., Carelli, R., Kelly, R., & Zannatha, J. M. I. (2006). Coordinated control of mobile

robots based on artiﬁcial vision. International Journal of Computers Communications &

Control, 1 (2), 85–94. https://doi.org/10.15837/ijccc.2006.2.2288

Tan, H., Chen, C., Luo, X., Zhang, J., Seibold, C., Yang, K., & Stiefelhagen, R. (2021). Fly-

ing guide dog: Walkable path discovery for the visually impaired utilizing drones and

transformer-based semantic segmentation. 2021 IEEE International Conference on Robotics

and Biomimetics (ROBIO), 1123–1128. https://doi.org/10.1109/robio54168.2021.9739520

Technologies, U. (n.d.). Unity - Manual: Unity User Manual 2021.3 (LTS). https://docs.unity3d.

com/Manual/UnityManual.html

Truchetet, F., Nicolier, F., & Laligant, O. (2001). Subpixel edge detection for dimensional control

by artiﬁcial vision. Journal of Electronic Imaging, 10 (1), 234–239. https://doi.org/10.1117/

1.1316089

van der Grinten, M. L., van Steveninck, J. d. R., Lozano, A., Pijnacker, L., Rueckauer, B., Roelf-

sema, P., van Gerven, M., van Wezel, R., Guclu, U., & Gucluturk, Y. (2022). Biologically

plausible phosphene simulation for the diﬀerentiable optimization of visual cortical prosthe-

ses. bioRxiv.

Vergnieux, V., Mac´e, M. J.-M., & Jouﬀrais, C. (2017). Simpliﬁcation of visual rendering in simulated

prosthetic vision facilitates navigation. Artiﬁcial organs, 41 (9), 852–861. https://doi.org/

10.1111/aor.12868

World Blind Union. (n.d.). https://worldblindunion.org/

R. E. Lucas Master Thesis

Zhao, Y., Kupferstein, E., Tal, D., & Azenkot, S. (2018). ”It Looks Beautiful but Scary”. https:

//doi.org/10.1145/3234695.3236359

R. E. Lucas Master Thesis

Appendix A

Table 3: MANOVA on the eﬀects of number of phosphenes, type of information-reduction method

and trial type on performance

Eﬀect Wilks’ Λ DF p

intercept 0.093 1 <.001**

phospheneNumber 0.820 2 <.001**

trialName 0.831 3 <.001**

informationMethod 0.827 2 <.001**

phospheneNumber * trialName 0.972 6 <.001**

phospheneNumber * informationMethod 0.972 4 0.501

informationMethod * trialName 0.937 6 0.002**

phospheneNumber * trialName * informationMethod 0.968 12 0.998

Residuals 608

Total 645

Note. *p < .05. **p < .01

Table 4: ANOVA on the eﬀects of number of phosphenes, type of information-reduction method

and trial type on the number of obstacle bumps

Source SS df MS F p

informationMethod 12596.358 2 6298.179 21.433 < .001**

phosphenesNumber 21473.956 2 10736.978 36.539 < .001**

phosphenesNumber * informationMethod 1779.799 4 444.950 1.514 0.196

trialName 20095.731 3 6698.577 22.796 < .001**

Residuals 185711.700 632 293.848

Total 241657.544 645

Note. *p < .05. **p < .01

Table 5: ANOVA on the eﬀects of number of phosphenes, type of information-reduction method

and trial type on the number of road bumps

Source SS df MS F p

phosphenesNumber 239.252 2 119.626 6.382 0.002**

informationMethod 2028.023 2 1014.011 54.095 < .001**

phosphenesNumber * informationMethod 143.704 4 35.926 1.917 0.106

trialName 385.644 3 128.548 6.858 < .001**

Residuals 11846.910 632 18.745

Total 14643.533 645

Note. *p < .05. **p < .01

R. E. Lucas Master Thesis

Table 6: ANOVA on the eﬀects of number of phosphenes, type of information-processing method

and trial type on the time needed to complete a trial

Source SS df MS F p

phosphenesNumber 98469.503 2 49234.752 66.094 < .001**

informationMethod 30490.624 2 15245.312 20.466 < .001**

phosphenesNumber * informationMethod 5451.210 4 1362.802 1.829 0.121

trialName 40690.413 3 13563.471 18.208 < .001

Residuals 470787.557 632 744.917

Total 645889.307 645

Note. *p < .05. **p < .01

Table 7: Post hoc test for obstacles with diﬀerent number of phosphenes

Mean Diﬀerence SE t p

tukey

64 625 10.246 1.658 6.182 < .001**

1000 13.621 1.658 8.218 < .001**

625 1000 3.375 1.649 2.049 0.102

Note. *p < .05. **p < .01

Table 8: Post hoc test for road with diﬀerent number of phosphenes

Mean Diﬀerence SE t p

tukey

64 625 0.898 0.419 2.145 0.097

1000 1.486 0.419 3.549 0.001 ∗ ∗

625 1000 0.588 0.417 1.411 0.476

Note. *p < .05. **p < .01

Table 9: Post hoc test for time with diﬀerent number of phosphenes

Mean Diﬀerence SE t p

tukey

64 625 21.825 2.639 8.270 < .001**

1000 29.215 2.639 11.070 < .001**

625 1000 7.390 2.636 2.814 0.014*

Note. *p < .05. **p < .01

Table 10: Post hoc test for obstacles with diﬀerent methods

Mean Diﬀerence SE t p

tukey

SS SW 7.251 1.658 4.375 <.001**

WP 10.631 1.658 6.413 <.001**

SW WP 3.380 1.649 2.049 0.102

Note. *p < .05. **p < .01

R. E. Lucas Master Thesis

Table 11: Post hoc test for road with diﬀerent methods

Mean Diﬀerence SE t p

tukey

SS SW 2.713 0.419 6.480 < .001**

WP 4.310 0.419 10.295 < .001**

SW WP 1.597 0.417 3.834 < .001**

Note. *p < .05. **p < .01

Table 12: Post hoc test for time with diﬀerent methods

Mean Diﬀerence SE t p

tukey

SS SW 10.374 2.636 3.931 < .001**

WP 16.737 2.639 6.342 < .001**

SW WP 6.363 2.636 2.432 0.041*

Note. *p < .05. **p < .01

Table 13: ANOVA on the eﬀects of number of phosphenes, type of information-reduction method

and trial type on the subjective experience

Source SS df MS F p

phosphenesNumber 258.415 2 129.208 40.317 < .001**

informationMethod 154.948 2 77.474 24.175 < .001**

phosphenesNumber * informationMethod 8.071 4 2.018 0.630 0.641

Residuals 4098.891 1279 3.205

Total 4520.325 1287

Note. *p < .05. **p < .01

Table 14: Post hoc test for phosphenes on subjective experience

Mean Diﬀerence SE t p

tukey

64 625 0.824 0.123 6.724 < .001**

1000 1.047 0.123 8.537 < .001**

625 1000 0.222 0.122 1.824 0.162

Note. *p < .05. **p < .01

Table 15: Post hoc test for method on subjective experience

Mean Diﬀerence SE t p

tukey

SS SW 0.650 0.123 5.303 < .001**

WP 0.789 0.121 6.515 < .001**

SW WP 0.139 0.123 1.130 0.496

Note. *p < .05. **p < .01

R. E. Lucas Master Thesis

Appendix B

We also performed Bayesian ANOVAs to test the previous results obtained with a classical

statistical paradigm. These tests showed that our previous results were robust.

The tables of these tests can be found here.

Table 16: Analysis of Eﬀects - obstacle bumps

Eﬀects P(incl) P(excl) P(incl—data) P(excl—data) BF

incl

phosphenesN 0.737 0.263 1.000 1.542 × 10

−13

2.316 × 10

+12

informationM 0.737 0.263 1.000 1.076 × 10

−7

3.321 × 10

trial 0.737 0.263 1.000 6.876 × 10

−12

5.194 × 10

+10

phosphenesN * informationM 0.316 0.684 0.072 0.928 0.169

phosphenesN * trial 0.316 0.684 0.033 0.967 0.074

informationM * trial 0.316 0.684 0.125 0.875 0.310

phosphenesN * informationM * trial 0.053 0.947 9.064 × 10

−6

1.000 1.632 × 10

−4

Table 17: Analysis of Eﬀects - road bumps

Eﬀects P(incl) P(excl) P(incl—data) P(excl—data) BF

incl

informationM 0.737 0.263 1.000 2.220 × 10

−15

1.608 × 10

+14

trialName 0.737 0.263 0.992 0.008 46.483

informationM * trialN 0.316 0.684 0.507 0.493 2.227

phosphenesN 0.737 0.263 0.892 0.108 2.956

informationM * phosphenesN 0.316 0.684 0.124 0.876 0.307

trial * phosphenesN 0.316 0.684 0.004 0.996 0.008

informationM * trial * phosphenesN 0.053 0.947 7.924 × 10

−7

1.000 1.426 × 10

−5

Table 18: Analysis of Eﬀects - time

Eﬀects P(incl) P(excl) P(incl—data) P(excl—data) BF

incl

trial 0.737 0.263 1.000 2.971 × 10

−9

1.202 × 10

informationM 0.737 0.263 1.000 1.917 × 10

−7

1.863 × 10

trialN * informationM 0.316 0.684 0.007 0.993 0.016

phosphenesN 0.737 0.263 1.000 3.331 × 10

−15

1.072 × 10

+14

trial * phosphenesN 0.316 0.684 0.004 0.996 0.008

informationM * phosphenesN 0.316 0.684 0.097 0.903 0.233

trial * informationM * phosphenesN 0.053 0.947 1.843 × 10

−8

1.000 3.318 × 10

−7