Evaluation of Augmented Reality Frameworks for Android Development


Published on

Master defense at Graz University of Technology, 2014

Published in: Education
  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide
  • We have identified 5 target criteria.
    The target images are printed both in color and grayscale. The tests determine which frameworks also support grayscale target images. The testing time for a color target image can be compared to the time needed to recognize its grayscale version and consequently it can be observed if tracking is affected.
    The target images are printed out with different contrast values. The tests determine how high/low the contrast ratio can reach in order for the target image to still be recognized. The contrast levels are from left to right, -50 and 0 on the top row and 50 and 100 on the second row.
    The target images are printed in different sizes, 5, 10, 15 and 20cm. All tests are performed placing the target images at a fixed distance from the camera, starting from small sizes and progressively increasing the size. These tests determine the minimum and maximum size a target image can have such that they are still recognized by the framework.
    The original target image is shrunk either vertically or horizontally by reducing the height, respectively the width to one third (1/3), a half (1/2) and two thirds (2/3).
    The target images are printed on different materials. Tracking is influenced by the material on which a target image is printed as it might be more difficult to recognize the target from a glass than a sheet of paper. A number of different materials are tested, including:
    Paper, normal printing paper.
    Glass, for bottle surfaces.
    Plasticized paper for restaurant menus.
    Glossy photo paper, for the mirroring effect.
  • Here we mention the constant flicker, a visible motion blur, and the ability to deal with fast moves such that the virtual content is not lost.
    Registration, which is the accurate alignment between the real world and the virtual object, is an important factor to overcome, as well as the capability to occlude the virtual object when necessary to create the feeling that the virtual object belongs in the real world, and is part of the scene.
  • The considered features include face tracking, text detection, flash and usage of the front camera.
    Moreover, the ability to display the virtual content even when the target image is not in the line of sight anymore, known as extended tracking, is a useful feature.
    Likewise, the possibility to track more than one target image simultaneously falls under the usability category.
  • For the purpose of evaluating the frameworks, we have implemented a test app that has 6 main views as shown here: Home view, Frameworks view, Criteria view, Scenarios view, Test view and Results view.
    Criteria and criteria categories can be added to the list of predefined criteria.
    Use cases can be defined by adding weighted criteria into a context.
    The frameworks can be compared against each other given a constraint or a set of constraints. Furthermore, an overview for each framework is available, displaying the average testing times per framework for each criteria category.
  • Two types of tests can be performed by the Test App: active tests and offline tests.
    The active tests answer the question “how much time does a framework need to recognize a target image under a specific constraint?”
    By actively testing, it is meant that a framework is chosen, its camera view opens and an online criterion is selected for testing.
    When the testing environment is set, the user starts the test timer by pressing the Start button. When the virtual content is superimposed into the real world, the testing time is saved as the time needed to detect and recognize the target image.
    We determine if the performance and usability criteria are supported by observing the behaviour of the frameworks in a neutral context, on the default target images (no special light conditions). Some feature support, such as face tracking and text detection, are determined by finding the information in the framework documentation.
  • On each criterion, all frameworks were subsequently evaluated more than once. What can be seen in these graphs represent the average testing times obtained and now I will present the most interesting results.
    The light intensity tests revealed a weak ARToolKit tracker when dealing with sudden change in light conditions or a semidark environment. Each of the considered four sequences for testing light intensities instigate environmental issues.
    Angles of 45degrees are generally overcome. However, issues are raised for ARToolKit which has the highest average times and is the only framework that does not recognize the target image at a 45 degree angle to both the left and the right.
    Metaio is the only framework that reaches the 240cm distance marker, granted, it needs over 10 seconds to recognize the target.
    A 40% visible area of the target is a reasonable constraint for almost every framework. Nevertheless, some frameworks work with even less visibility. Metaio produced an interesting outcome: if the entire target image is visible within a camera frame, but a percentage of it is obscured, metaio will not recognize the target. If however, only the visible area is within the camera frame and the obscured part not, metaio will recognize the target from a 40% visible area up.
    Testing the frameworks on the default size target from various distances brought some interesting results. Up close, 10cm away from the target image, the target is in most cases not completely visible. However, ARToolKit and ARLab recognized the targets. This is surprising because ARToolKit needs a 80% visible area of the target, while Vuforia only requires 10% and still could not recognize the target up close. From these findings it is concluded that the size of the target image and the distance to it are not strongly correlated.
    Another surprising result was registered by Vuforia which can detect a target image with a 200% level of noise. Most of the frameworks can still overcome a 70%, some even 90% noise level, but 200% is an achievement.
  • Grayscale targets and tests with various contrast ratios have been successfull.
    Four different target sizes have also been tested and it was determined that not all frameworks support different size varieties.
    Some of the frameworks come with the recommendation to not modify the target’s aspect ratio. For this purpose, it was tested what would be the effect of "shrinking" the target image. As a result, it is concluded that metaio is the only robust framework to various aspect ratios and D’Fusion is highly adaptable.
  • We have defined 6 scenarios for illustrating our findings.
    Consider an indoors app for visualizing large objects like furniture, home furnishings, appliances. A target image is placed in the spot where the real object would be and the virtual object is displayed on top of that target.
    Assume the scenario of a magazine app. An image in the magazine represents a target image for the magazine app. By hovering over with a smartphone the virtual content is revealed to the user.
    Suppose a company comes up with a new marketing idea and uses Augmented Reality for posters creatively located on the side panels of bus shelters. A user can be either at the bus stop looking at the poster or notice it from inside a moving bus and tries to decipher its message.
    Some supermarkets today use Augmented Reality to promote their products. Passing by a shelf with various products, the customer can point the camera to the boxes, select a product and initiate a game. For example burst some bubbles, such that when a predefined number of busted bubbles is reached, the customer wins a discount for that product.
    An app exists today for translating text from English into Spanish and several other languages. Imagine you are visiting a foreign country, you do not know the language and need help to get around. This app can be used to translate signs, window ads, menus.
    A number of applications already exist for virtually trying on products such as glasses, hats or wrist watches and order them online. Take glasses for example. By using such an app, the customer can browse through the catalogue, choose a pair of glasses, switch between available colors and try them on. The glasses are projected on the image of the face looking back from the phone.
  • Here is the definition of the interior design scenario. A couple of constraints that must be considered include:
    • how far or close the user can be from the target (distance).
    • having different perspectives for a better outlook of the virtual object (viewpoint).
    • occluding the virtual object by real objects if necessary (occlusion).
    • a correct alignment between the real scene and the virtual object (registration).
    • using more than one target for a better placement of the virtual object, at the right place and at the right scale (multi-targets).
    • seeing the virtual object even when the target is not visible anymore (extended tracking).
  • The magazine app is defined by most of the target criteria:
    • gray colored images (grayscale).
    • strong contrasts images (contrast ratio).
    • small size images (size).
    • images printed either on normal paper or glossy paper (material).
    • recognize images given that they are not completely visible (visibility).
  • Evaluation of Augmented Reality Frameworks for Android Development

    1. 1.  S C I E N C E  T E C H N O L O G Y  P A S S I O N u www.iicm.tugraz.at Evaluation of Augmented Reality Frameworks for Android Development Iulia Marneanu July 3rd, 2014
    2. 2.  Which is currently the best open framework for developing an Augmented Reality mobile application? July 3rd, 2014 Iulia Marneanu 2
    3. 3.  Agenda • Motivation • Augmented Reality • AR Frameworks • Criteria • Test App • Scenarios & findings July 3rd, 2014 Iulia Marneanu 3
    4. 4.  Motivation • Evolaris Next Level GmbH • growing in popularity on mobile devices [1] • 200 million mobile AR users worldwide by 2018 [2] • over 2.5 billion mobile AR apps by 2017 [3] • recent job market growth in Android development [4] July 3rd, 2014 Iulia Marneanu 4
    5. 5.  Augmented Reality July 3rd, 2014 Iulia Marneanu 5 [5]
    6. 6.  • Marker-based tracking • GPS tracking • Markerless tracking Augmented Reality. Tracking Techniques July 3rd, 2014 Iulia Marneanu 6
    7. 7.  AR Frameworks Framework Development Availability ARLab Spain (2006) Demo ARToolKit USA (2001) Academic License D‘Fusion France (1999) Watermark Vuforia Austria (2011) Free catchoom Spain (2011) Demo metaio Germany (2003) Watermark July 3rd, 2014 Iulia Marneanu 7
    8. 8.  Criteria • Environmental criteria • Target criteria • Performance criteria • Usability criteria July 3rd, 2014 Iulia Marneanu 8
    9. 9.  Criteria. Environmental Criteria July 3rd, 2014 Iulia Marneanu 9 Environmental Criteria Tests Light intensity - Natural light - Direct light - Sudden change - Mirroring Viewpoint 45° angle perspectives Distance From 10 cm distance between testing device and target. Visibility From 10% visible area. Background Dark versus bright contrasts. Noise From 10% noisy target image.
    10. 10.  Target Criteria Tests Grayscale Default size target image in grayscale. Contrast Ratio Contrast value set to -50. Size Four sizes: 5cm, 10cm, 15cm and 20cm. Aspect Ratio Vertically or horizontally Material - Glass - Laminated - Glossy photo paper Criteria. Target Criteria July 3rd, 2014 Iulia Marneanu 10
    11. 11.  Criteria. Performance Criteria July 3rd, 2014 Iulia Marneanu 11 Frameworks Flicker Motion Blur Fast Moves Registration Occlusion ARLab no no not supp. not supp. not supp. ARToolKit min no partially supported not supp. not supp. D‘Fusion no no supported supported not supp. Vuforia min no supported supported not supp. catchoom no yes supported supported not supp. metaio min no not supp. supported not supp.
    12. 12.  Criteria. Usability Criteria July 3rd, 2014 Iulia Marneanu 12 Frameworks Face Tracking Text Detection Flash Camera Front Camera Extended Tracking Multi Targets ARLab yes yes ARToolKit yes yes D‘Fusion yes yes yes Vuforia no yes yes yes yes catchoom no no yes yes metaio yes yes no yes yes yes
    13. 13.  Test App July 3rd, 2014 Iulia Marneanu 13 Home Frameworks Criteria Scenarios Test Result
    14. 14.  Test App. Tests • Active tests „how much time does it need to…?“ • Online criteria • Record the testing times • Offline tests „is it supported or not?“ • Offline criteria • Determine if a feature is supported or not July 3rd, 2014 Iulia Marneanu 14
    15. 15.  Test App. Environmental Criteria Results July 3rd, 2014 Iulia Marneanu 15
    16. 16.  Test App. Target Criteria Results July 3rd, 2014 Iulia Marneanu 16
    17. 17.  Scenarios • Interior Design • Magazine App • Bus Shelter • Supermarket Promotions • Tourist Translator • mCommerce July 3rd, 2014 Iulia Marneanu 17
    18. 18.  Scenarios. Interior Design Environmental Criteria Weight Light Intensity 2 Viewpoint 5 Visibility 2 Distance 4 July 3rd, 2014 Iulia Marneanu 18 Target Criteria Weight Size 3 Performance Criteria Weight Registration 5 Occlusion 4 Usability Criteria Weight Multi-targets 3 Extended Tracking 4
    19. 19.  Scenarios. Magazine App Environmental Criteria Weight Light Intensity 5 Viewpoint 3 Visibility 4 Noise 3 Distance 1 July 3rd, 2014 Iulia Marneanu 19 Target Criteria Weight Grayscale 4 Contrast Ratio 1 Size 2 Aspect Ratio 2 Material 3 Performance Criteria Weight Registration 1 Flicker 1 Usability Criteria Weight Multi-targets 3 Text Detection 2
    20. 20.  Scenarios. Bus Shelter Environmental Criteria Weight Light Intensity 5 Viewpoint 5 Visibility 4 Noise 5 Background 4 Distance 3 July 3rd, 2014 Iulia Marneanu 20 Target Criteria Weight Size 2 Aspect Ratio 2 Material 3 Performance Criteria Weight Fast Moves 4 Usability Weight Text Detection 1 Extended Tracking 4 Flash 1
    21. 21.  Scenarios. Supermarket Promotions Environmental Criteria Weight Light Intensity 1 Viewpoint 4 Visibility 5 Noise 2 Distance 2 July 3rd, 2014 Iulia Marneanu 21 Target Criteria Weight Grayscale 3 Size 3 Aspect Ratio 3 Material 4 Performance Criteria Weight Motion Blur 4 Usability Criteria Weight Multi-targets 5 Text Detection 1 Flash 1
    22. 22.  Scenarios. Tourist Translator Environmental Criteria Weight Light Intensity 4 Viewpoint 4 Visibility 4 Noise 2 Background 1 Distance 3 July 3rd, 2014 Iulia Marneanu 22 Target Criteria Weight Contrast Ratio 5 Size 3 Material 2 Performance Criteria Weight Flicker 5 Usability Criteria Weight Text Detection 5 Flash 1
    23. 23.  Scenarios. mCommerce Environmental Criteria Weight Light Intensity 5 Viewpoint 4 Visibility 4 Background 1 Distance 3 July 3rd, 2014 Iulia Marneanu 23 Performance Criteria Weight Flicker 2 Fast Moves 3 Usability Criteria Weight Face Tracking 5 Front Camera 4
    24. 24.  No AR framework is better than another, each has its advantages and disadvantages. In some circumstances, given a set of constraints, a framework can outperform others. July 3rd, 2014 Iulia Marneanu 24
    25. 25.  References [1] TechNavio - Infiniti Research Ltd. Global augmented reality market 2014- 2018. A market research report. [2] Juniper Research. Mobile augmented reality: smartphones, tablets and smart glasses 2013-2018. A market research report. [3] Juniper Research. Over 2.5 Billion Mobile Augmented Reality Apps to Be Installed Per Annum by 2017. A market research report. [4] Jennifer Lynn. iOS vs. OS: Current Job Market Is Growing Faster For Android Developers Than Apple. [5] https://www.youtube.com/watch?v=Go9rf9GmYpM July 3rd, 2014 Iulia Marneanu 25
    26. 26.  Thank you for your attention! July 3rd, 2014 Iulia Marneanu 26
    27. 27.  July 3rd, 2014 Iulia Marneanu 27