DocTr: A Unifying Framework for Tracking PhysicalDocuments and Organisational Structures

Sandra Trullemans, Ayrton Vercruysse and Beat SignerWeb & Information Systems Engineering Lab

Vrije Universiteit BrusselPleinlaan 2, 1050 Brussels, Belgium{strullem,ayvercru,bsigner}@vub.ac.be

ABSTRACTDespite major advancements in digital document manage-ment, paper documents still play an important role in ourdaily work and are often used in combination with digitaldocuments and services. Over the last two decades, we haveseen a number of augmented reality solutions helping usersin managing their paper documents in office settings. How-ever, since data is mainly managed at the application layer,the use of multiple document tracking setups results in frag-mented and inconsistent tracking data. Furthermore, exist-ing tracking solutions often focus on the tracking of paperdocuments in organisational structures such as folders or fil-ing cabinets without taking into account the flow of docu-ments across these organisational structures. We present theDocument Tracking (DocTr) framework for unifying existingdocument tracking setups and managing document metadataacross organisational structures. The DocTr framework hasbeen implemented based on a user-centric requirements anal-ysis and simplifies the development of interactive computingsystems for personal cross-media information management.

ACM Classification KeywordsH.5.m. Information Interfaces and Presentation (e.g. HCI):Miscellaneous

Author KeywordsDocument tracking; personal information management;organisational structures.

INTRODUCTIONDespite the predictions of the paperless office in the 1970s,we are still using paper documents in our daily work [29].In many office environments digital as well as physical docu-ments form a part of our workflows. It has further been shownthat frequent tablet users often print digital documents in or-der to facilitate the task at hand [4]. Recent studies highlightthat besides the personal preferences for paper documents,

Permission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor profit or commercial advantage and that copies bear this notice and the full cita-tion on the first page. Copyrights for components of this work owned by others thanACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re-publish, to post on servers or to redistribute to lists, requires prior specific permissionand/or a fee. Request permissions from Permissions@acm.org.EICS’16, June 21–24, 2016, Brussels, Belgium© 2016 ACM. ISBN 978-1-4503-4322-0/16/06$15.00DOI: http://dx.doi.org/10.1145/2933242.2933254

companies in various business sectors do not get rid of paperdocuments due to productivity or law reasons [11].

Since the early 1990s, applications and technologies havebeen developed which enable the tracking of paper docu-ments in a user’s working environment. Thereby, paperdocuments are tagged with two-dimensional tags such asbarcodes or electronic markers like Radio Frequency Iden-tification (RFID) tags. As an alternative to the explicittagging of documents, computer vision techniques can beused to identify paper documents. These document track-ing technologies enabled various research projects such as theDigitalDesk [35] project where paper is integrated with an in-teractive desk. Users can further be supported in organisingand re-finding their documents via so-called Personal Infor-mation Management (PIM) solutions. There are a number ofPIM solutions for managing paper documents which usuallyconsist of a paper document tracking component and somedigital search functionality. In addition to the absolute posi-tion of a document in space, some of these applications alsotrack to which organisational structure (e.g. pile or folder) adocument belongs to. Note that existing solutions are oftenlimited to track documents within a specific organisationalstructure in order to provide users access to the correspond-ing digital document version. For example, the SIFT algo-rithm has been applied for tracking interactions with docu-ment piles [17] whereas in [28] the physical folders in a filingcabinet have been augmented based on visual tags and digitalpen and paper technology.

In this paper, we use the term tracking setup for a particu-lar setup including hardware as well as software componentswhich enable the tracking of paper documents within a spe-cific area based on tagging or some of the other technologiesmentioned above. For example, a tracking setup might con-sist of a camera which is mounted above the desk in com-bination with some computer vision software or it can be abookcase where individual books have been RFID tagged.While typical office environments consist of multiple organ-isational structures and paper documents naturally flow be-tween these structures, existing solutions focus on trackingdocuments within single organisational structures based ondedicated technologies. Furthermore, any data managementis typically delegated to the application layer, leading to frag-mented and inconsistent tracking data when multiple trackingsetups are used.

We start by investigating the existing body of work in thedomain of document tracking solutions. The detailed anal-ysis of existing solutions in combination with an exploratoryuser study led to a number of user-centric as well as tech-nical design requirements which are discussed in detail andform the basis of our unifying document tracking framework.We then describe the DocTr framework for integrating differ-ent document tracking setups and managing document flowsacross organisational structures. Furthermore, we illustratehow DocTr unifies data management for third-party applica-tions based on the OC2 PIM framework [33]. A description ofhow third-party applications can profit from the DocTr frame-work is followed by a preliminary evaluation of the frame-work and some concluding remarks.

BACKGROUNDGiven the fact that paper documents still play an importantrole in our daily work, various research efforts have been un-dertaken to integrate paper documents and physical storageartefacts (e.g. a filing cabinet) with digital information sys-tems. Interactive tabletop surfaces are for example used incombination with cameras above the desk for augmenting pa-per documents with digital functionality. Thereby, paper doc-uments are often tagged with a two-dimensional (2D) code asseen in DocuDesk [7]. A number of rapid prototyping frame-works have been developed to ease the use of tags for ob-ject recognition. For instance, the reacTIVision [14] frame-work detects fiducial markers (i.e. 2D tags with a specificpattern) within a well-defined surface area and notifies regis-tered client applications about detected tags and their relativeorientation. Note that the reacTIVision framework is widelyused in tabletop and augmented desk applications includingObjectTop [16] and iCon [5].

The tagging of paper documents interrupts the user’s regulartasks and therefore various computer vision-based documentidentification solutions have been developed as alternativesto tagging. However, due to changing environmental factorssuch as the lighting conditions, computer vision-based solu-tions are often more vulnerable to errors than tag-based so-lutions. Already in their seminal DigitalDesk system, Well-ner [35] used feature extraction on captured document imagesin order to identify which document had been placed on aninteractive desk. Nowadays, more advanced solutions, suchas the Scale Invariant Feature Transformation (SIFT) [21]algorithm proposed by Lowe or the Speed Up Robust Fea-tures (SURF) [1] algorithm, are used for document recog-nition. The SIFT algorithm has, for example, been appliedin FACT [20] to enable fine-grained cross-media interactionsbetween paper documents and a laptop. In addition to paperdocuments also other artefacts can be recognised. The workof Matsushita et al. [23] tracks books in a bookshelf based onthe SURF algorithm. Since paper documents often containa number of unique text blocks, Optical Character Recogni-tion (OCR) techniques can be applied to the text in order toidentify a document and any written text can be transformedinto a digital form [34].

The aforementioned tracking techniques have also been usedfor tracking paper documents in organisational structures

such as piles or filing cabinets. A number of applications aug-ment existing physical organisational structures with track-ing functionality in order to provide a digital representationof a specific organisational structure. The work of Lawrieand Rus [19] describes an early application providing a dig-ital representation of a physical filing cabinet. Paper docu-ments in piles on a desk can be tracked by using computervision techniques [17, 25]. Besides these traditional organisa-tional structures, people also use other artefacts such as boxesand drawers for storing documents or other physical objects.DrawerFinder [18] provides tracking support for such organ-isational structures by tagging drawers or boxes and monitor-ing a user’s interaction with a storage shelf containing thesetagged boxes.

While the previously mentioned applications monitor onespecific sort of organisational structure (e.g. a pile), dur-ing their lifecycle documents naturally move between dif-ferent organisational structures. Therefore, there is a needto track documents across organisational structures. TheHuman-Centered Workplace system monitors printed docu-ments across organisationla structures by augmenting themwith a Quick Response (QR) code at printing time [6].PaperSpace [31] uses a similar tracking approach but addsinteractive printed buttons to paper documents. Users can,for example, point to a printed button to request the digitalversion of a document. While these applications track paperdocuments across organisational structures, they do not keeptrack of the internal state of these organisational structures.For example, while the Human-Centered Workplace systemmonitors paper documents in an office environment by us-ing multiple cameras and provides applications the absoluteposition of tracked documents, it does not model the individ-ual organisational structures (e.g. piles). In the context of thepresented DocTr framework, SOPHYA [12] is the most rel-evant solution. The SOPHYA framework provides a way totrack ordered collections of artefacts such as folders whichhave been augmented with electronic circuits. In addition,artefacts can be enhanced with re-finding user interfaces suchas an LED strip in a bookshelf which lights up when a usersearches for the corresponding artefact [10]. Although theSOPHYA framework is extensible to support various kinds oforganisational structures, it is necessary that all tracked arte-facts are augmented with electronic circuits which might notbe an ideal solution for every office setting.

In an office environment, users have different kinds of organ-isational structures and paper documents can move betweeninstances of these organisational structures. Such office en-vironments might therefore require multiple of the trackingsolutions presented in this section. In the remaining part ofthis paper we introduce DocTr, a framework that unifies ex-isting tracking solutions and manages a document’s lifecycleacross different organisational structures.

DESIGN REQUIREMENTSThe development of the DocTr framework has been basedon a set of user-centric and technical design requirements.These requirements have been derived from an exploratory

user study as well as an investigation of the existing body ofdescriptive PIM research.

Exploratory User StudyIn order to define a number of user-centric design require-ments, we have conducted an exploratory user study. We in-terviewed eleven participants aged between 23 and 56. Allparticipants used a significant amount of paper documentsduring their working activities. Furthermore, they had differ-ent professions such as secretary, social worker, middle man-ager and managing director.

Since most participants were not familiar with paper docu-ment tracking solutions, we first informed them about the ex-isting body of related work. We introduced DocuDesk andDigitalDesk, tracking books in bookshelves, DrawerFinder,the Human-Centered Workspace and SOPHYA to the partic-ipants by showing them what they could do with these sys-tems based on screenshots of the corresponding publications.Our introduction focussed on HCI and no details about tech-nical aspects were provided. Furthermore, in order to min-imise bias we explained the possible interactions in an ob-jective and neutral manner. In a next step, we conducteda semi-structured interview to investigate issues with exist-ing solutions and to explore opportunities for improvement.The main questions among others were: “For which reasonswould you consider to use one of the current solutions?”, “Doyou see shortcomings or disadvantages of these solutions?”,“How do you see the future of tracking technologies?”, “Whatshould be the minimal functionality of a tracking solution?”and “Can you describe how the ideal tracking solution couldbe integrated in your working environment?”. Note that sincethe focus is on determining generic requirements, the givenquestions did not refer to any specific tracking systems.

Based on these interviews, we identified some shortcomingsof existing solutions. First, participants mentioned that us-ing a single technology is not desired due to the fact thattheir document organisation is very dynamic. Moreover, pa-per documents are often moved around. At the same time,they did not want to be overwhelmed by multiple isolated so-lutions (e.g. different applications for each tracking setup).Finally, our participants would not consider to invest in ex-isting solutions due to a lack of integration with their currentapplications such as the File Explorer, Microsoft Word andEvernote.

User-centric Design RequirementsBased on the interviews and related descriptive PIM research,we derived the following user-centric design requirements inorder to improve current paper document tracking solutions.

R1: Categories of Organisational StructuresIn descriptive PIM research, organisational structures can beclassified in three categories including files (e.g. filing cabi-nets), piles and mixtures (e.g. unordered documents in a lettertray) [22, 32]. Previous research argues that during re-findingactivities in these different categories of organisational struc-tures users rely on different cues (i.e. context, space or time).In a file structure, users will mostly use the context cue suchas orienteering between documents which were used in the

same task. In contrast, when a user searches information in apile, they often try to recall the position of the required doc-ument in the pile (e.g. at the bottom or top) [36, 32]. In ad-dition, our interviews show that users prefer a semantic de-scription of a document’s position such as “the document isin the ring binder with the label Bills”. For seven out of theeleven participants, it would already be enough to just men-tion the category of the organisational structure containing therequested document, as mentioned by a participant: “In mostcases, this would trigger my spatial cue and then my memorywill take it over. I do not think that I will often need the exactlocation”. In order to enable the design of PIM applicationswhich provide support for one or multiple re-finding cues andto inform users about the categories of organisational struc-tures, we have to integrate tracking support for the three cat-egories of organisational structures.

R2: Flow of Documents and Organisational StructuresAll participants identified the limited support for tracking pa-per documents across organisational structures as the mainreason for the lack of acceptance of current tracking tech-nologies. They do not see the benefits of installing differ-ent tracking setups with their individual applications. In con-trast, they would see this as a burden as stated by a partici-pant “I think people are becoming crazy with all those differ-ent apps everywhere. If all these setups would again requiremore apps, I first have to find the right app. By then I havefound my paper sheet!”. Besides the flow of paper documentsacross organisational structures, participants also highlightedthat they often move the organisational structures themselvessuch as when creating a new pile out of two existing pilesor reorganising ring binders. Therefore, a tracking solutionshould also foresee the tracking of the organisational struc-tures themselves. In addition to the results of our interview,the flow of documents has also been investigated in descrip-tive PIM studies. It has been shown that paper documents cancontain cold, warm or hot information [29]. So-called colddocuments are often archived in a filing structure whereaspiles seem to contain mostly hot documents (i.e. documentswhich are often used) [29, 32]. In addition, it has been shownthat a user’s personal information space contains a lot of colddocuments which are rarely accessed or for which the usersimply forgot that they exist [36]. By tracking paper docu-ments across organisational structures and managing the flowof organisational structures themselves, users can be offered aunified solution and PIM systems can, for example, integratethe history of a paper document’s flow and make users awareof rarely accessed paper documents.

R3: Custom MetadataAs mentioned before, most participants prefer a semantic de-scription of a paper document’s position. Therefore, track-ing setups should be able to provide end users some seman-tic document metadata. In addition, the interviews revealedthat users frequently would like to provide their own meta-data about specific paper documents. Note that this custommetadata does not necessarily consist of positional data butcan contain arbitrary annotations. Users prefer to add custommetadata as observed in current practise where they annotatedocuments with contextual information [29]. One cause can

be seen in the fact that when applying the filing strategy, theoverall context of a document is lost [13]. A flexible doc-ument tracking solution should therefore offer a mechanismfor managing various kinds of positional as well as customdocument metadata.

R4: Integration with Third-party ApplicationsEight of the eleven participants preferred to have the doc-ument tracking metadata available in their frequently usedtools. A number of participants mentioned that they wouldfind it convenient to have a digital representation of pa-per documents and their metadata in applications such asEvernote where these paper documents could then be in-cluded in digital notes. Other participants proposed to justmake the data available in the File Explorer in order that theycould organise paper documents together with their digitalmedia. These findings are in line with previous work. It hasbeen observed that users hold on to their familiar tools whenit comes down to PIM [3]. In addition, SOPHYA which isclosely related to the presented work, aims for a decoupling oftracking technologies and third-part applications [12]. There-fore, it should be possible to integrate a tracking solution withthird-party applications and these applications should haveeasy access to any document tracking metadata.

Technical Design RequirementsBesides the design requirements resulting from our ex-ploratory user study, we identified a number of technical re-quirements for offering the functionality previously describedin the user-centric design requirements.

R5: Unlimited Tracking SetupsWhile existing solutions focus on particular recognition tech-niques, developers should be enabled to easily combine ex-isting tracking solutions in their applications. Since an inter-active office environment is highly user and application de-pendent (i.e. users have different storage artefacts and customoffice configurations), there needs to be a separation of con-cerns between the used tracking setups and the applicationsmaking use of paper document metadata. Moreover, specifictracking setups should only be responsible for detecting a pa-per document’s metadata such as its position while the storageof this metadata should be managed by a central repositorywhich is accessible by third-party applications. Note that thistechnical requirement will enable the tracking of paper doc-uments across organisational structures (R2) and support aneasy integration with third-party applications (R4).

R6: Integration with a PIM FrameworkIn the PIM research field, frameworks such as Gnowsis [26],HayStack [15] and MyLifeBits [8] have been developed fordocument management in personal information spaces. Com-monly, these systems manage personal documents in a cen-tral repository and provide metadata about digital documentswhich can be used in organising and re-finding activities. Along-term study by Sauermann and Heim [27] highlightedthe promising future of such PIM frameworks. Nevertheless,most PIM frameworks only support digital media. By also in-tegrating paper documents with these frameworks, users canbe offered a unified cross-media PIM solution (R4). Second,by making use of a PIM framework, any paper document

Document Management

Pend

ing

Pool

Dat

a M

odel

Iden

tific

atio

n M

echa

nism

OC

2

GUI

REST Input Interface

REST Output Interface

Third-party Applications

DocTr

Del

egat

or

Tracking Setups

Figure 1: DocTr architecture overview

metadata can be unified with the metadata of digital docu-ments. The integration with a PIM solution would furthermake it possible to let users and applications store their owncustom metadata (R3) since most PIM frameworks offer thisfunctionality. In addition, it is commonly accepted that thereis a need for document management in office settings and thata PIM solution can be used for managing information aboutthe three categories of organisational structures [32] (R1). Ul-timately, the integration of a PIM framework will lead to amore unified personal information space.

R7: Managing Unique IdentifiersDue to the possibility of combining tracking setups with theirown unique document identifiers, a general document track-ing solution has to provide a mechanism for managing andmapping the different unique identifiers that might be as-signed to a single paper document by the different trackingsolutions. The challenge lies in the fact that unique identi-fiers can take different formats. For example, a documentrecognised by the SIFT algorithm will be identified by a setof extracted features while tag-based recognition frameworksuse an integer or string as identifier. By supporting multipleunique identifiers for a single paper document or organisa-tional structure, the flow of items across different organisa-tion structures and tracking setups can be managed (R2).

DOCTR FRAMEWORKWe now present the DocTr framework which addresses theseven design requirements that have been defined in the previ-ous section. As illustrated in Figure 1, the framework forms amiddleware between multiple tracking setups and third-partyapplications that want to make use of a document’s trackedmetadata. The different tracking setups can be as simple as

Figure 2: DocTr exploratory graphical user interface

listening to reacTIVision for detected tags or they might con-sist of complex software to extract paper document featuresvia the SIFT algorithm.

There is a clear separation of concerns between the respon-sibilities of different tracking setups and the DocTr frame-work. The responsibility of a tracking setup is to determinethe unique identifier of a traced paper document. In addi-tion, a tracking setup is in charge of providing DocTr thedigital representation of a monitored paper document. Sub-sequently, the DocTr framework makes these digital repre-sentations available to third-party applications. It is up to aspecific tracking setup to define how this digital representa-tion looks like. A representation could, for example, consistof an image of the paper document or a custom icon. In caseno digital representation has been assigned to a monitored pa-per document, third-party applications will receive a defaulticon by the DocTr framework which is shown in DocTr’sgraphical user interface shown in Figure 2. When a track-ing setup detects a paper document, it sends the document’sunique identifier together with some derived metadata such asthe documents position to DocTr. DocTr then checks whetherthe paper document is moving between organisational struc-tures or whether it is a new paper document. In the case thatthis process does not lead to a clear result, the user is askedfor clarification. Users can interact with DocTr via the ex-ploratory graphical user interface shown in Figure 2 as well asthe sidebar shown in Figure 3 where users can provide feed-back about documents with open tracking issues.

The implementation of DocTr is based on a client-server ar-chitecture with the DocTr server using the component-basedsoftware architecture shown in Figure 1. Tracking setupscan communicate with DocTr via the REST Input Interfacewhile third-party applications can use the REST Output In-terface to query metadata about tracked paper documents andorganisational structures. The Document Management com-ponent is responsible for the data management of trackeddocuments and organisational structures while the Identifi-cation Mechanism is in charge of comparing unique docu-

ment identifiers. The Document Management component in-cludes the OC2 PIM framework [33] in order to fulfil thesecond technical requirement (R6). We have extended theOC2 data model to support different categories of organi-sational structures (R1) and enable the storage of custommetadata (R3). When DocTr cannot automatically determinewhether a tracked entity has been moved, the correspondingdocument is forwarded to the Pending Pool component. Paperdocuments and organisational structures that have been added

Figure 3: DocTr sidebar

to the Pending Pool wait for further user feedback such as aconfirmation that a document has been moved if DocTr is notsure about its position or if a document has been placed in anon-trackable area. The communication between the differ-ent components is coordinated by the Delegator component.

The DocTr framework has been implemented in Java and usesa GlassFish1 server together with the Atmosphere2 frame-work providing an implementation of the JAX-RS recom-mendation for RESTful web services. In the following, wefurther elaborate on the individual DocTr components andtheir implementation.

DocTr Data Model and PIM IntegrationWe have chosen to use the OC2 PIM framework [33] foroffering the necessary PIM functionality. The OC2 frame-work enables the linking of digital and physical documentsvia bidirectional navigational and structural links as definedby the RSL hypermedia metamodel [30]. Navigational linksare used to express navigational paths between digital andphysical documents while structural links provide a way toexpress structures within documents. Furthermore, OC2 al-lows us to store document metadata and define how relevanta document is in a given context or task. Last but not least,the OC2 PIM solution also offers some user management.

In order to support organisational structures (R1), custommetadata (R3) and multiple unique identifiers (R6), wehave slightly extended the OC2 data model. An Entity-Relationship (ER) representation of the resulting extendeddata model with the shaded entities that have already been de-fined in OC2 is shown in Figure 4. In the OC2 model, objectscan represent digital or physical objects as indicated with adisjoint constraint between the subclasses of Object. Thisallows us to create an instance of a physical object for eachpaper document. As mentioned before, a tracking setup is re-sponsible to provide unique identifiers for their tracked paperdocuments and documents can be tracked by various track-ing setups. Therefore, an Object can be associated withmultiple tracking setups where each association between anobject and a specific tracking setup can have multiple track-ing properties. In this way, we can store custom metadataof a specific paper document for a specific tracking setup.Since the unique identifier of a paper document depends onthe tracking setup monitoring the document, it is stored as atracking property. The one-to-many constraints on the rela-tionship between tracking properties and tracking setups en-ables the reuse of unique document identifiers across trackingsetups. For example, a paper document might be identifiedby a fiducial marker which can of course be used by two ormore tracking setups. Furthermore, it is necessary to managemultiple unique identifiers of paper documents with differentformats (R7). Therefore, we introduced the Comparatorentity. For a specific format, a comparator component willcompare a given detected identifier against a set of uniqueidentifiers. In order to allow the identification mechanism tocompare unique identifiers, each tracking setup has to specifythe appropriate comparator for its tracked paper documents.1https://glassfish.java.net/2https://github.com/Atmosphere/

Note that the tracking functionality is modelled at the level ofan Object. In the future this allows us to reuse the DocTrdata model for digital media such as the tracking of digitaldocuments across cloud spaces.

Digital Object

Physical Object

(0,*)

(1,*)

Object

hasLog

Tracking Setup

Loghas

Tracking

Tracking Property

Comparator has

(1,*)

(1,*)

(1,*)

(1,*)(0,*)

Organisatio-nalLink

hasSource

hasTarget

Structure

Structural Link

PileFile Mixture

d

d

hasElements

(1,1)

(1,1)

(0,*) (0,*)

(0,*)

(0,*)

Organisatio-nalStructure

d

d

Figure 4: DocTr data model with highlighted OC2 entities

Besides offering this tracking setup functionality, we haveto enable the storage of custom metadata (R3). Therefore,an object can be assigned some Log instances. A log takesthe format of (TimeStamp, Transactions) where atransaction consists of an action and its value. Note that incontrast to other tracking applications, a timestamp can havemultiple transactions. This enables applications to decom-pose an action in sub-actions. For example, applications canstore one log entry when a paper document is removed froman organisational structure with the action ‘remove’ havinga value ‘pile1’ and at the same time with an action ‘byUser’with the name of the user who moved the paper document.

Organisational structures are stored as an organisational linkwith an Object entity as source and an ordered set ofObject entities as target. The source object represents theorganisational structure such as a pile while the targets of theorganisational link represent the documents included in theorganisational structure. Note that by modelling the organ-isational structure itself as an Object entity, the model’stracking extensions can also be applied to organisational

structures. This is necessary since organisational structuresare also dynamic in their position and a ring binder can,for example, be moved from a shelf to a desk. In addi-tion, organisational structures have to be categorised in oneof the three organisational structure types (R1). Since theOC2 data model already provides the notion of structures,we have introduced the OrganisationalStructureas a subclass of the Structure entity. In its turn, theOrganisationalStructure entity is subclassified intothe File, Pile and Mixture organisational structures.The disjoint constraint indicates that an organisational struc-ture can only take the form of one type of organisational struc-ture. Finally, an organisational link has to be associated withan organisational structure.

Identification MechanismThe identification mechanism is responsible for finding aPhysical Object instance based on a detected uniqueidentifier. Since physical objects can have multiple uniqueidentifiers in different formats, we have introduced theComparator components. As mentioned before, eachtracking setup indicates the Comparator instance to beused when processing received unique identifiers. Currently,DocTr supports comparators for SIFT features, String com-parators based on the Levenshtein distance as well as perfectmatching for Strings and Integers. From our experience, theseare the most commonly used formats in existing tracking se-tups. However, custom comparators can be provided by im-plementing the compare() method of the Comparatorinterface shown in Listing 1. The method has two parame-ters. The first parameter is a map with Physical Objectas key entry and a list of Objects representing an object’sunique identifier values. The second parameter is the receivedunique identifier (i.e. wrapped in an Object instance) whichhas to be compared against the values provided in the map in-stance. When the received unique identifier can be mapped toa physical object’s identifier, the physical object is returned.

Listing 1: Comparator interfacei n t e r f a c e Compara tor

p u b l i c P h y s i c a l O b j e c t compare (Map<P h y s i c a l O b j e c t , L i s t<Objec t >>, O b j e c t ) ;

When a tracking setup sends the unique identifier of a de-tected paper document to DocTr, the Delegator fetchesthe corresponding Comparator and queries the data modelfor all tracking properties where the value can be comparedwith the corresponding comparator. This set of unique iden-tifiers is then compared to the received unique identifier bythe Comparator instance. If there is a match, the physicalobject is returned to the Delegatorwhich requests the doc-ument management component to update the data model withthe received metadata. Otherwise, the Delegator forwardsthe unique identifier to the document management componentwhere it is added in the pending pool.

Pending PoolThe Pending Pool component manages paper documentswhich could not be resolved by the identification mechanism.

This can, for example, happen when a paper document thatis uniquely identified by a tag is moved to another trackingsetup which uses SIFT features for identification. Similarly,organisational structures such as a ring binder could, for in-stance, be identified via an RFID tag when placed on a deskwhile computer vision techniques might be used to detect thering binder’s label when stored in a shelf. Furthermore, a usercan always throw away a document, give it to somebody elseor place it in an unmonitored area.

Our solution for addressing undetected documents is inspiredby context-aware applications which aim to involve the userin order to decrease the vulnerability to errors and increaseuser satisfaction [2]. We developed an end-user applicationwhere users can provide the necessary input to resolve issueswith documents in the pending pool. In addition, DocTr of-fers a exploratory graphical user interface to navigate throughthe tracked paper documents and explore their metadata. Theexploratory GUI presents the paper documents in a list show-ing the digital representation of the paper documents. In casethat the tracking setup did not provide a digital representa-tion, a default icon is used. Users can right click on a doc-ument in order to request a document’s details as shown inFigure 2. The details panel includes the document’s track-ing properties such as the unique identifiers (uris) but alsoto which organisational structure (structureId) the doc-ument belongs to. Furthermore, users can add annotationsto the document in the note section. When a document isadded to the pending pool, the user will get a notification viathe system tray. The system tray allows quick access to theDocTr sidebar shown in Figure 3. Unidentified paper docu-ments are added to the sidebar’s upper part. A user can clickon the green plus button to indicate that the document is newor use the search button to open the exploratory GUI. Afterselecting a document, a user can inform DocTr that it is thesame document as the one shown in the sidebar by selectingthe Same As context menu. The pending pool will then addthe new unique identifier to the selected document. When apaper document is removed from an organisational structureand cannot be allocated to another structure, the document isadded to the lower part of the sidebar. A user can choose todelete a document (bin) or to add it to another organisationalstructure. In case that a document should be assigned to an-other organisational structure, a popup is shown with all or-ganisational structures containing non-identified documentsas well as an area where a user can indicate whether a doc-ument has been placed outside of a monitoring area or givento someone else. This metadata will be stored in the docu-ment’s log. Depending on a user’s actions, the pending poolwill notify the Delegator component to proceed with the re-quired actions such as adding the physical object instance as atarget to the organisational link at a certain index of the givenorganisational structure.

HOW TO USE DOCTROn the server side, we have foreseen two REST interfaceswhich can be used by client applications to communicatewith DocTr as illustrated earlier in Figure 1. We distinguishbetween physical document tracking client applications and

client applications making use of the tracking metadata inorder to ease the development of interactive office environ-ments. The REST Input Interface is meant for tracking setupswhich add information about documents in their monitoredarea while the REST Output Interface is used by third-partyapplications intending to use metadata about tracked paperdocuments and organisational structures.

Input Interface for Tracking SetupsThe REST Input Interface includes endpoints for creatingnew documents as well as organisational structures and forupdating metadata values. In addition, it can also be usedto interact with the identification mechanism for comparingunique identifiers across tracking setups. First, a trackingsetup has to register itself and define which comparator isrequired for its unique identifiers. In the simplest case, atracking setup only monitors paper documents without anyorganisational structures. When a document’s unique iden-tifier (i.e. tag number or SIFT features) is detected in thissimple case, the tracking setup sends the unique identifier to-gether with additional log entries to DocTr by calling the cor-responding endpoint for updating a physical document. TheDelegator component requests the identification mechanismfor the given physical object and checks whether the detecteddocument is already waiting for allocation in the pendingpool. Next, the Delegator updates the data model with thenew knowledge. When a tracking setup monitors an organ-isational structure, it has to create this structure in DocTr inan initialisation phase and keep the returned identifier of theinitialised organisational structure. When detecting a changeto the organisational structure, the tracking setup has to notifyDocTr by sending a JSON string including the identifier of theorganisational structure, the index where the document has tobe added and the derived unique identifier of the tracked phys-ical document. When a document is removed from an organ-isational structure or monitoring area, the tracking setup hasto call the remove endpoint with the identifier of the organi-sational structure and the document identifier. As previouslymentioned, the document will be added to the pending pool ifno other tracking setup recognised the removed document.

Output Interface for Third-party ApplicationsWhile the REST Input Interface is used by tracking setups,the REST Output Interface can be used by third-party appli-cations to query DocTr’s data model as well as to listen toupdates to the data model. The interface provides endpointsfor querying the data model entities and exploring documentsof the pending pool. Furthermore, applications can establisha WebSocket connection to be notified when a data model en-tity has changed its value. A JSON message is sent over theWebSocket connection including the identifier of the entityand the new values such as a new location.

INTEGRATION WITH EXISTING SOLUTIONSIn order to illustrate how DocTr can be used in practice, weprovide three use cases in different application domains. Notethat these use cases have not yet been implemented but ratherillustrate the potential and future directions in the develop-ment of interactive computing systems based on DocTr.

Integration with Office EnvironmentsNormally, users have different storage artefacts and config-urations of their office environments which require differenttracking technologies. Based on the fact that DocTr helpsintegrating different tracking setups, an interactive office ap-plication no longer has to be based on a proprietary platformfor managing different tracking setups. In practice, we could,for example, use the SOPHYA [12] framework for trackingphysical folders in bookcases. In addition, a user might havea monitored area around their laptop on a desk which iden-tifies paper documents based on fiducial markers as done inDocuDesk [7]. In their daily activities, a user might move apaper document from an augmented SOPHYA folder to theirworkplace around the laptop. DocuDesk and SOPHYA wouldonly have to deal with four DocTr endpoints including theregistration, the initialisation of organisational structures, aswell as the adding and removal of endpoints. Of course, atracking setup can always make more extensive use of DocTr.It could, for example, use DocTr to store log entries everytime a user accesses a document or store so-called cross-media organisational structures such as a pile that containsphysical as well as digital documents.

Integration with Third-party ApplicationsDocTr can also easily be integrated with third-party appli-cations. For example, an Evernote extension can be devel-oped which listens to DocTr for log updates of paper docu-ments which a user has used in Evernote notes. In addition,Evernote can request the digital representation of paper doc-uments in order to allow users to integrate paper documentsin their Evernote notes. Similarly, the digital representationof a paper document can, for example, be used in the FileExplorer in order that users can organise digital and physicaldocuments in a unified way. This functionality can be im-plemented by invoking a WebSocket connection. DocTr willpublish new updates of the data model in the form of JSONmessages on this WebSocket channel. Third-party applica-tions only have to process the received messages which con-tain the identifier of the document or organisational structureas well as any additional metadata.

Integration with Context-aware FrameworksIn the digital information space, metadata about digital doc-uments such as the creation time and access frequencies areprovided by most operating systems. Third-party applicationscan easily access this information. For example, recent re-search in context-aware applications such as CAAD [24] andPassage [9] use this metadata and other data extracted by textmining to determine a user’s current activity. Due to the si-multaneous use of digital and physical documents in our dailyworkflows, we can improve these activity recognition appli-cations by providing metadata about paper documents in asimilar way as done for their digital counterparts.

EVALUATIONThe accuracy of DocTr depends on the accuracy of the usedtracking setups and the implementation of the comparatorsand an evaluation of DocTr as a middleware is therefore onlypossible by integrating it in larger applications with a signifi-cant amount of different tracking setups. Therefore, we have

1 2

3

4

T1 T2 T3

T4aT4b

Figure 5: Evaluation setup

chosen for a preliminary evaluation which investigates howwell DocTr can handle error-prone tracking setups such asthe SIFT algorithm and how the behaviour changes over timewhen multiple unique identifiers are given to a document.The results show that even with error-prone tracking, the ac-curacy improves by letting users provide feedback about sim-ilar documents in the pending pool.

We have implemented the four tracking setups shown in Fig-ure 5 in order to drive the evaluation. The first two trackingsetups 1© and 2© monitor piles based on the SIFT algorithm.The third tracking setup 3© tracks the adding and removalof paper documents from ring binders on a desk surface. Fi-nally, the last tracking setup 4©monitors a bookcase with ringbinders. Note that the presented results are focussing on thefeatures of DocTr and the accuracy of the document trackingdepends on the used document tracking setups.

MethodologyThe four tracking setups have been used to determine the ac-curacy of DocTr when moving paper documents across or-ganisational structures with the same or different documentidentifiers formats as well as for the movement of the organ-isational structures themselves. Note that in the beginning ofthe evaluation, documents have been added to the pile moni-tored by the first tracking setup 1©.

T1: Moving a paper document between organisational struc-tures with the same document identifier formatTracking setups 1© and 2© both track physical documentsbased on the SIFT algorithm. We have used the SIFT Key-point Detector Java library3 to extract the SIFT features ofdocuments. Furthermore, as shown in Figure 5 the two track-ing setups use different cameras. Note that each task Ti con-sists of a subtask T R

i which removes a document from a track-ing setup and a subtask T A

i adding the same document to an-other tracking setup. In this first task T1 of our evaluation,

3http://www.cs.ubc.ca/˜lowe/keypoints

documents are removed from the pile of tracking setup 1©via T R

1 and added to the pile of tracking setup 2© via T A1 .

T2: Moving a paper document between organisational struc-tures with different document identifier formatsIn order to evaluate the fact that a physical document canhave more than one unique identifier across tracking setups,T R

2 moves the documents from the pile in tracking setup 2©and T A

2 adds them to a ring binder monitored by trackingsetup 3©. Tracking setup 3© consists of RFID tagged ringbinders, an RFID reader placed underneath the desk and acamera installed above the desk. When physical documentsare moved from the pile of tracking setup 2©, we added a fidu-cial marker to them. In the case a ring binder is placed withinthe range of the RFID reader, the application starts listening tothe reacTIVision framework in order to detect fiducial mark-ers. Finally, if any document is detected, the application re-quests DocTr to add the detected document to the ring binderorganisational structure.

T3: Moving an organisational structureSince organisational structures can also be tracked by theDocTr framework, in task T3 the ring binder was removedfrom the desk via T R

3 and added to the bookcase via T A3 . Each

shelf of the bookcase contains an RFID reader and beforeplacing the ring binder on a shelf, it has to be scanned bythe associated RFID reader.

T4: Moving documents back to the first tracking setupFinally, we moved the documents of the ring binder back tothe pile of tracking setup 1©. As part of task T4, T R

4a firstmoved the ring binder from the bookcase shelf and T A

4a addedit to tracking setup 3©. The documents were then removedfrom the ring binder via T R

4b and added to the pile by T A4b.

The procedure consisting of task T1 to T4 was repeated fivetimes in order to present more accurate results. Furthermore,a total of 55 printed research papers from various conferencesand journals were used. We decided to use research papersdue to their high similarity when processed by computer vi-

Figure 6: Percentage of positively identified documents

sion techniques which represented a challenge the used doc-ument identification techniques.

ResultsAs described in the methodology, in each task physical docu-ments are first removed from an organisational structure andthen added to another organisational structure and the twosubtasks are monitored by different tracking setups. The re-sults of the five iterations for the four described tasks T1 to T4are shown in Figure 6. For each iteration and task Ti, we mea-sured the amount of positively identified documents when re-moving them via subtask T R

i from on organisational structureand adding them to another organisational structure via T A

i .

When having a look at the results of T R1 , T A

1 and T R2 , we can

observe that in the first two iterations the identification mech-anism could not accurately match the given unique identifiersbased on SIFT. This might be caused by the error pronenessof the used SIFT algorithm and the small threshold valuedefined in the SIFT comparator. In each iteration, the newunique SIFT identifier is manually added to non-recogniseddocuments via the DocTr sidebar. As a result, the matchingaccuracy improved over the five iterations without reaching100%. This might be caused by the fact that the two SIFTtracking setups both define their SIFT feature matrices whichmight slightly diverge from each other. Note that the amountof unmatched physical documents were caused by the inac-curacy of the used SIFT algorithm in the two tracking setupsand by using research papers which looked very similar. Itis known that the SIFT algorithm is not ideal for identifyingsimilar documents [17] but DocTr can help to improve theaccuracy of error-prone tracking setups via user feedback.

In T2 documents are removed from the pile of trackingsetup 2© via subtask T R

2 and added to the ring binder 3© viasubtask T A

2 . During the first iteration of T2, documents wereaugmented with a tag before being added to the ring binderand the DocTr sidebar has been used to allocate the tag num-ber to the document. As a consequence, none of the docu-ments could be identified during this first iteration. The re-sults of T A

2 further show that documents are always positivelyidentified after the first iteration. Furthermore, removing thering binder from the desk via subtask T R

3 , adding it to the

shelf via T A3 and placing it back to the desk via subtasks T R

4aand T A

4a is always positively tracked. Finally, the removal ofdocuments from the ring binder via T R

4b has also been trackedaccurately. Nevertheless, when adding the documents back tothe first pile of tracking setup setup 1© via T A

4b, due to SIFT-related issues we could observe the same behaviour as for theresults of T1 and T2. Note that in the first iteration of T A

4b nodocuments were identified by the SIFT comparator since tagshave been added to the documents in the first iteration of T A

2 .

DISCUSSION AND FUTURE WORKWhile physical document tracking solutions are recentlygaining attention, existing solutions show some limitations.They often only track physical documents in one specifictype of organisational structure such as piles or physical fold-ers. In addition, the metadata management of tracked doc-uments is forwarded to the application layer which leads tofragmentation and inconsistency when multiple tracking se-tups are used. In order to overcome these issues, we pre-sented the DocTr framework which unifies existing trackingsetups and provides extensive data management possibilities.With DocTr documents can be tracked across organisationalstructures regardless of the used tracking setups. In addition,we also support the traceability of the organisational struc-tures themselves which is normally not supported by existingsolutions. The DocTr framework currently offers a simplegraphical user interface for exploring tracked documents. Weplan to develop a more advanced GUI where users can or-ganise their physical documents by also defining associationsbetween them. Furthermore, in a long-term in-context eval-uation we plan to investigate the usability of DocTr and gaininsights about the trade-off between the effort it takes to main-tain the pending pool and not up-to-date tracking metadata.Since physical documents are often used together with digi-tal media, in the near future DocTr might be extended to alsodeal with the management of a user’s digital documents. Thiswould ultimately enable the unified organisation of physicaland digital documents through the development of so-calledcross-media PIM applications. Finally, there is also potentialto further improve the identification mechanism when multi-ple versions of an identifier are used by removing the versionsthat have not been matched for a long time.

CONCLUSIONWe have presented DocTr, a unifying framework for track-ing physical documents and organisational structures acrossdifferent tracking setups. DocTr deals with data provided byvarious tracking setups and makes this data available to third-party applications. The exploratory graphical user interfaceand DocTr sidebar further enable end users to explore trackeddocuments and provide feedback for documents that have notbe recognised correctly. Our general DocTr framework of-fers an ideal platform for future research on interactive com-puting systems including interactive office environments andpersonal cross-media information management in particular.

ACKNOWLEDEGMENTSThe research of Sandra Trullemans is funded by the Agencyfor Innovation by Science and Technology in Flanders (IWT).

REFERENCES1. Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. 2006.

SURF: Speeded up Robust Features. In Proceedings ofEuropean Conference on Computer Vision (ECCV2006). Springer Berlin Heidelberg, 404–417. DOI:http://dx.doi.org/10.1007/11744023_32

2. Victoria Bellotti and Keith Edwards. 2001. Intelligibilityand Accountability: Human Considerations inContext-Aware Systems. Human-Computer Interaction16, 2–4 (2001), 193–212. DOI:http://dx.doi.org/10.1207/S15327051HCI16234_05

3. Ofer Bergman, Ruth Beyth-Marom, Rafi Nachmias, NoaGradovitch, and Steve Whittaker. 2008. ImprovedSearch Engines and Navigation Preference in PersonalInformation Management. Transactions on InformationSystems 26, 4 (2008), 20. DOI:http://dx.doi.org/10.1145/1402256.1402259

4. Paris Buttfield-Addison, Christopher Lueg, Leonie Ellis,and Jon Manning. 2012. “Everything Goes Into or Outof the iPad”: The iPad, Information Scraps and PersonalInformation Management. In Proceedings of AustralianComputer-Human Interaction Conference (OzCHI2012). ACM, 61–67. DOI:http://dx.doi.org/10.1145/2414536.2414546

5. Kai-Yin Cheng, Rong-Hao Liang, Bing-Yu Chen,Rung-Huei Laing, and Sy-Yen Kuo. 2010. iCon:Utilizing Everyday Objects as Additional, Auxiliary andInstant Tabletop Controllers. In Proceedings of theSIGCHI Conference on Human Factors in ComputingSystems (CHI 2010). ACM, 1155–1164. DOI:http://dx.doi.org/10.1145/1753326.1753499

6. Amay Dighe and Annika Hinze. 2012. Human-CentredWorkplace: Re-finding Physical Documents in an OfficeEnvironment. In Proceedings of the Annual Conferenceof the New Zealand Chapter of the ACM Special InterestGroup on Computer-Human Interaction (CHINZ 2012).ACM, 9–16. DOI:http://dx.doi.org/10.1145/2379256.2379259

7. Katherine M. Everitt, Meredith Ringel Morris,A.J. Bernheim Brush, and Andrew D. Wilson. 2008.DocuDesk: An Interactive Surface for Creating andRehydrating Many-to-Many Linkages Among Paper andDigital Documents. In Proceedings of IEEEInternational Workshop on Horizontal InteractiveHuman Computer Systems (TABLETOP 2008). IEEE,25–28. DOI:http://dx.doi.org/10.1109/TABLETOP.2008.4660179

8. Jim Gemmell, Gordon Bell, Roger Lueder, StevenDrucker, and Curtis Wong. 2002. MyLifeBits: Fulfillingthe Memex Vision. In Proceedings of the 10th ACMinternational Conference on Multimedia(MULTIMEDIA 2002). ACM, 235–238. DOI:http://dx.doi.org/10.1145/641007.641053

9. Karl Gyllstrom. 2009. Passages Through Time:Chronicling Users’ Information Interaction History byRecording When and What They Read. In Proceedings

of the International Conference on Intelligent UserInterfaces (IUI 2009). ACM, 147–156. DOI:http://dx.doi.org/10.1145/1502650.1502673

10. Matthew Jervis and Masood Masoodian. 2013.Visualization of Physical Library Shelves to FacilitateCollection Management and Retrieval. In Proceedingsof the Symposium on Engineering InteractiveComputing Systems (EICS 2013). ACM, 133–138.DOI:http://dx.doi.org/10.1145/2494603.2480321

11. Matthew Jervis and Masood Masoodian. 2014. How DoPeople Attempt to Integrate the Management of TheirPaper and Electronic Documents? Aslib Journal ofInformation Management 66, 2 (2014), 134–155. DOI:http://dx.doi.org/10.1108/AJIM-01-2013-0007

12. Matthew G. Jervis and Masood Masoodian. 2010.SOPHYA: A System for Digital Management ofOrdered Physical Document Collections. In Proceedingsof the Fourth International Conference on Tangible,Embedded, and Embodied Interaction (TEI 2010).ACM, 33–40. DOI:http://dx.doi.org/10.1145/1709886.1709894

13. William Jones and Jaime Teevan. 2007. PersonalInformation Management. University of WashingtonPress.

14. Martin Kaltenbrunner and Ross Bencina. 2007.reacTIVision: A Computer-Vision Framework forTable-Based Tangible Interaction. In Proceedings of theInternational Conference on Tangible and EmbeddedInteraction (TEI 2007). ACM, 69–74. DOI:http://dx.doi.org/10.1145/1226969.1226983

15. David Karger, Karun Bakshi, David Huynh, DennisQuan, and Vineet Sinha. 2005. Haystack: ACustomizable General-Purpose InformationManagement Tool for End Users of SemistructuredData. In Proceedings of the 1st Biennial Conference onInnovative Data Systems Research (CIDR 2005). 13–26.

16. Mohammadreza Khalilbeigi, Jurgen Steimle, JanRiemann, Niloofar Dezfuli, Max Muhlhauser, andJames D. Hollan. 2013. ObjecTop: Occlusion Awarenessof Physical Objects on Interactive Tabletops. InProceedings of the ACM international Conference onInteractive Tabletops and Surfaces (ITS 2013). ACM,255–264. DOI:http://dx.doi.org/10.1145/2512349.2512806

17. Jiwon Kim, Steven M. Seitz, and Maneesh Agrawala.2004. Video-Based Document Tracking: Unifying YourPhysical and Electronic Desktops. In Proceedings of the17th Annual ACM Symposium on User InterfaceSoftware and Technology (UIST 2004). ACM, 99–107.DOI:http://dx.doi.org/10.1145/1029632.1029650

18. Mizuho Komatsuzaki, Koji Tsukada, and Itiro Siio.2011. DrawerFinder: Finding Items in Storage BoxesUsing Pictures and Visual Markers. In Proceedings ofthe International Conference on Intelligent UserInterfaces (IUI 2011). ACM, 363–366. DOI:http://dx.doi.org/10.1145/1943403.1943466

19. Dawn Lawrie and Daniel Rus. 1999. A Self-OrganizedFile Cabinet. In Proceedings of the InternationalConference on Information and KnowledgeManagement (CIKM 1999). ACM, 499–506. DOI:http://dx.doi.org/10.1145/319950.320056

20. Chunyuan Liao, Hao Tang, Qiong Liu, Patrick Chiu, andFrancine Chen. 2010. FACT: Fine-Grained Cross-MediaInteraction With Documents Via a Portable HybridPaper-Laptop Interface. In Proceedings of theInternational Conference on Multimedia (MM 2010).ACM, 361–370. DOI:http://dx.doi.org/10.1145/1873951.1874001

21. David Lowe. 1999. Object Recognition From LocalScale-Invariant Features. In Proceedings of theInternational Conference on Computer Vision (ICCV1999). IEEE, 1150–1157. DOI:http://dx.doi.org/10.1109/ICCV.1999.790410

22. Thomas W. Malone. 1983. How People Organize TheirDesks? Implications for the Design of OfficeInformation Systems. ACM Transactions on OfficeInformation Systems (TOIS) 1, 1 (1983), 99–112.

23. Kazuhiro Matsushita, Daisuke Iwai, and Kosuke Sato.2011. Interactive Bookshelf Surface for In Situ BookSearching And Storing Support. In Proceedings of theAugmented Human International Conference (AH2011). ACM. DOI:http://dx.doi.org/10.1145/1959826.1959828

24. Tye Rattenbury and John Canny. 2007. CAAD: anAutomatic Task Support System. In Proceedings of theConference on Human Factors in Computing Systems(CHI 2007). ACM, 687–696. DOI:http://dx.doi.org/10.1145/142750.143007

25. Jan Riemann, Mohammadreza Khalilbeigi, NiloofarDezfuli, and Max Muhlhauser. 2015. StackTop: HybridPhysical-Digital Stacking on Interactive Tabletops. InProceedings of the Conference on Human Factors inComputing Systems (CHI EA 2015). ACM, 1127–1132.DOI:http://dx.doi.org/10.1145/2702613.2732742

26. Leo Sauermann, Gunnar Aastrand Grimnes, MalteKiesel, Christiaan Fluit, Heiko Maus, Dominik DanishHeim, DominHeim, Benjamin Horak, and AndreasDengel. 2006. Semantic Desktop 2.0: The GnowsisExperience. In Proceedings of the InternationalSemantic Web Conference (ISWC 2006). Springer BerlinHeidelberg, 887–900. DOI:http://dx.doi.org/10.1007/11926078_64

27. Leo Sauermann and Dominik Heim. 2008. EvaluatingLong-Term Use of the Gnowsis Semantic Desktop forPIM. In Proceedings of the International Semantic WebConference (ISWC 2008). Springer Berlin Heidelberg,

467–482. DOI:http://dx.doi.org/10.1007/978-3-540-88564-1_30

28. Thomas Seifried, Jervis Matthew, Michael Haller,Masood Masoodian, Masoodian, and Nicolas Villar.2008. Integration of Virtual and Real DocumentOrganization. In Proceedings of the InternationalConference on Tangible and Embedded Interaction (TEI2008). ACM, 81–88. DOI:http://dx.doi.org/10.1145/1347390.1347410

29. Abigail Sellen and Richard Harper. 2001. The Myth OfThe Paperless Office. MIT Press.

30. Beat Signer and Moira C. Norrie. 2007. As We MayLink: A General Metamodel for Hypermedia Systems.In Proceedings of the 26th International Conference onConceptual Modeling (ER 2007). Springer BerlinHeidelberg, 359–374. DOI:http://dx.doi.org/10.1007/978-3-540-75563-0_25

31. Jeff Smith, Jeremy Long, Tanya Lung, Mohd Anwar,and Sriram Subramanian. 2006. PaperSpace: A Systemfor Managing Digital and Paper Documents. InProceedings of the Conference on Human Factors inComputing Systems (CHI EA 2006). ACM, 1343–1348.DOI:http://dx.doi.org/10.1145/1125451.1125700

32. Sandra Trullemans and Beat Signer. 2014a. From UserNeeds to Opportunities in Personal InformationManagement: A Case Study on OrganisationalStrategies in Cross-Media Information Spaces. InProceedings of the 14th ACM/IEEE-CS Joint Conferenceon Digital Libraries (JCDL 2014). IEEE, 87–96. DOI:http://dx.doi.org/10.1109/JCDL.2014.6970154

33. Sandra Trullemans and Beat Signer. 2014b. Towards aConceptual Framework and Metamodel forContext-Aware Personal Cross-Media InformationManagement Systems. In Proceedings of theInternational Conference on Conceptual Modeling (ER2014). Springer International Publishing, 313–320.DOI:http://dx.doi.org/10.1007/978-3-319-12206-9_26

34. Pierre Wellner. 1991. The DigitalDesk Calculator:Tangible Manipulation on a Desk Top Display. InProceedings of the Symposium on User InterfaceSoftware and Technology (UIST 1991). ACM, 27–33.DOI:http://dx.doi.org/10.1145/120782.120785

35. Pierre Wellner. 1993. Interacting with Paper on theDigitalDesk. Commun. ACM 36, 7 (1993), 87–96. DOI:http://dx.doi.org/10.1145/159544.159630

36. Steve Whittaker and Julia Hirschberg. 2001. TheCharacter, Value and Management of Personal PaperArchives. Transactions on Computer-Human Interaction8, 2 (2001), 150–170.