Object category recognition is one of the most difficult problems in computer vision. It involves recognizing objects despite intra-class variations, viewpoint changes and background clutter. The goal of this thesis is to investigate robust invariant local image description and the selection of discriminative features. We show that class-discriminative scale-invariant features achieve excellent results for image-level categorization and object localization. We present solutions for two key problems: (i) we improve the quality of the image description based on a novel scale-invariant keypoint detection method and (ii) we integrate feature filtering techniques into our object models. Our novel scale-invariant detector is based on the idea of a ``maximally stable description'', i.e., the descriptor should be stable even in the presence of minor variations of the detector. The technique performs scale selection based on a region descriptor, here SIFT, and chooses regions for which this descriptor is maximally stable, i.e., the difference between descriptors extracted for consecutive scales reaches a minimum. This scale selection technique is applied to multi-scale Harris and Laplacian points. Experimental results evaluate the performance of our detector and show that it outperforms existing ones in the context of image matching, category and texture classification, as well as object localization. To construct object models based on discriminative features, we first cluster the scale-invariant descriptors and obtain a set of ``visual words''. We then estimate the discriminative information of these clusters based on different feature selection techniques---several of which are traditionally used in text retrieval. We discuss their properties---feature frequency, discriminative power, and redundancy---and analyze their performance in the context of image classification and object localization. We show that each task has different requirements, and indicate which selection techniques are the most appropriate. Experimental results for recognition on challenging large datasets demonstrate the performance of the approach.

.

Malgré les progrés importants réalisés ces dernières décennies, les systèmes de reconnaissance de la parole spontanée grand vocabulaire sont encore beaucoup trop lents pour être utilisés dans des contextes réels. Le travail effectué au cours de cette thèse s'attache alors à proposer des méthodes de réduction du temps de calcul de ces systèmes en vue de permettre leur utilisation dans de tels contextes.<br />Parmi les tâches pour lesquelles l'optimisation du temps de reconnaissance est possible, nous nous sommes concentrés sur le calcul des probabilités des modèles, tâche occupant à elle seule souvent plus de la moitié du temps global de traitement. Les distributions des modles acoustiques utilisés par les systèmes de reconnaissance sont généralement représentées par des mélanges de gaussiennes. Le calcul des probabilités d'émissions relatives est donc particulièrement lié au nombre de gaussiennes considérées dans ces mélanges. Notre travail s'est porté alors sur la proposition et l'évaluation de méthodes de sélection de gaussiennes.<br />Les méthodes de sélection de gaussiennes existantes sont basées sur la classification. Pour améliorer cette classification, nous avons proposé une approche de partitionnement hiérarchique basée sur la similarité entre les distributions gaussiennes. La distance de Kullback-Leibler symétrisée et pondérée a été introduite à cet effet. Suite à la classification des gaussiennes, un algorithme de sélection multi-niveaux est appliqué.<br />Les mélanges de distributions gaussiennes correspondent à des contextes d'apparition différents. Malheureusement, Les méthodes de sélection existantes ne prennent pas en compte ces différents contextes, puisque lorsque le nombre de gaussiennes par état est faible, toutes les distributions sont mélangées avant d' être regroupées. Ainsi, nous avons développé une méthode de sélection contextuelle des gaussiennes.<br />Les méthodes de sous-quantification vectorielle sont apparues comme une alternative aux approches basées sur la sélection des gaussiennes. En regroupant toutes les distributions, elles mélangent certains contextes. Ainsi, nous avons introduit une sous-quantification vectorielle contextuelle.

This item is not the definitive copy. Please use the following citation when referencing this material: Dillon, A. (1987) Knowledge acquisition and conceptual models: a cognitive analysis of the interface. In: D. Diaper and R.Winder (eds.) People and Computers III. Cambridge: Cambridge University Press, 371-379. Abstract: Understanding how users process the information available to them through the computer interface can greatly enhance our abilities to design usable systems. This paper details the results of a longitudinal psychological experiment investigating the effect of interface style on user performance, knowledge acquisition and conceptual model development. Through the use of standard performance measures, interactive error scoring and protocol analysis techniques it becomes possible to identify crucial psychological factors in successful human computer use. Results indicate that a distinction between "deep" and "shallow" knowledge of system functioning can be drawn where both types of user appear to interact identically with the machine although significant differences in their respective knowledge exists. The effect of these differences on user ability to perform under stress and transfer to similar systems is noted. Implications for the design of usable systems are discussed.

Winner of the 1997 Brown Publication Award from ECT. This item is not the definitive copy. Please use the following citation when referencing this material: McKnight, C., Dillon, A., and Richardson, J. (1996) User Centered Design of Hypertext and Hypermedia for Education. In: D. Jonassen (ed) Handbook of Research on Educational Communications and Technology. New York: Macmillan, 622-633 Abstract The chapter begins by describing the fundamental concepts of hypertext and gives a brief overview of the different philosophical perspectives manifest in the key figures of the field. It then considers the role of hypertext in learning, concluding from a review of empirical evaluations that many of the claims for hypertext have failed to be substantiated. It is argued that for a variety of conceptual and methodological reasons, it is extremely difficult to evaluate hypertext experimentally in an educational context. However, rather than simply abandon either hypertext or empirical evaluation, the chapter concludes by arguing for an empirically grounded, user centred approach to the design of hypertext based on a knowledge of the users, their tasks, the information space and the context in which the three interact.

Active Reading is an important part of a knowledge worker’s activities; it involves highlighting, writing notes, marking with symbols, etc., on a document. Many Active Reading applications have been designed in seeking to replicate the affordances of paper through digital-ink-based annotation tools. However, these applications require users to perform numerous steps to use various types of annotation tools, which impose an unnecessary cognitive load, distracting them from their reading tasks. In this thesis, we introduce GestAnnot, an Active Reading application for tablet computers that takes a fundamentally different approach of incorporating multi-touch gesture techniques for creating and manipulating annotations on an e-document, thus offering a flexible and easy- to-use annotation solution. Based on the literature review, we designed and developed GestAnnot and then performed lab and field evaluations of the software. In lab evaluation, GestAnnot performed better than one of the best existing annotation application in many aspects, including number of steps. The design was then refined based on the feedback received. The field evaluation of the improved design helped us to understand the performance of the application in the real world. We proposed a set of design guidelines through the feedback received from both evaluations, which any future Active Reading application could benefit from.

In this thesis we examine 360 degree curvilinear displays and their potential for the display of 3D information. We present two systems: a spherical display prototype and a cylindrical display. Our spherical display prototype used the lack of set user position and natural visibility constraints of a spherical display to create a 3D display system that treated the sphere as a volumetric display through the use of 360 degree motion parallax and stereoscopy. We evaluated these properties by examining how our use of stereoscopy and 360 degree motion parallax, might aid in the preservation of basic body orientation cues and in pose estimation tasks in a telepresence application with our final cylindrical display. Results suggest the combined presence of 360 degree motion parallax and stereoscopic cues within our cylindrical display significantly improved the accuracy with which participants were able to assess gaze and hand pointing cues, and to instruct others on 3D body poses. The inclusion of 360 degree motion parallax and stereoscopic cues also led to significant increases in the sense of social presence and telepresence reported by participants. / Thesis (Master, Computing) -- Queen's University, 2013-01-29 14:14:33.822

Trying to ‘learn’ the spatial layout of an environment is a common problem in certain application domains, such as military and emergency personnel training. Until recently this training was accomplished solely by providing maps and briefings of an environment. These methods, however, only provide topological (survey) knowledge of the environment, which pays little attention to the details of routes and landmarks that can only be acquired through the acquisition of procedural knowledge via navigation. Unlike previous experiments concerning spatial knowledge acquisition this work does not attempt to determine whether spatial knowledge acquisition is feasible. Such investigations have yielded a variety of results, yet all agree that spatial knowledge acquisition from a virtual environment is feasible if given enough exposure time. Accordingly, the aim of this thesis is to contribute towards a better understanding of how various individual differences and environmental factors impact the exposure time requirements needed for a person to acquire spatial knowledge from a virtual environment. Although the results of our investigation should be used with caution, we show that a one-size-fits-all situation is not possible when estimating the required exposure time that a user needs to acquire spatial knowledge. Moreover we provide a guide that allows a trainer to predict the required exposure time a person will require, by using the person's personal profile, and the environment's particular factors. In addition, we found that one of the tests we used during our investigation caused unnecessary frustration and confusion to our participants. This test is a standard way of finding a participant's orientation skill, and is commonly used in the area of spatial knowledge acquisition. Therefore, by recreating a new electronic version of the test and comparing the scores from both the new test and the old one our investigation showed that the scores on the new test were significantly higher for all participants. The training time was also lowered significantly. Our updated electronic version will be useful in future research. This test is available online at: www.newgztest.com.

Online software tutorials help a wide range of users acquire skills with complex software, but are not always easy to follow. For example, a tutorial might target users with a high skill level, or it might contain errors and omissions. Prior work has shown that user contributions, such as user comments, can add value to a tutorial. Building on this prior work, we investigate an approach to soliciting structured tutorial enhancements from tutorial readers. We illustrate this approach through a prototype called Antorial, and evaluate its impact on reader contributions through a multi-session evaluation with 13 participants. Our findings suggest that structuring tutorial contributions has positive impacts on both the number and type of reader contributions. Our findings also point to design considerations for systems that aim to support community-based tutorial refinement, and suggest promising directions for future research. / February 2017

The research question for this study is “What different design guidelines can be distilledfrom teachers in a distance education organization when they are tasked to imagine theinterface for synchronous video communication with students”. Previous research hasshown that live streaming technologies are a viable option for students but more research isneeded with the teacher perspective. By employing qualitative methods such as semi-structured interviews and holding a participatory design workshop with teachers atdistance learning facilitator in Sweden this study argues that an online learning platformshould be designed with social media aspects in mind, the ability to hold teacher-heldlectures on a regular basis, present different modes of participation in order to improveonline distance learning. The author suggests that more studies with the teacherperspective in mind should be conducted to further improve the understanding of teachers’needs in an online distance-learning environment.