|
University of South Florida |
|
To teach, to seek, and to learn |
|
Here are some of the major research thrusts. Other research directions include: vision-based navigation, stereo, light source calibration, edge detection and scale space. |
|
Sign languages are complex, abstract linguistic systems, with their own grammars. We are concerned with building automated algorithms that can take sign language video of and recognize the signs performed. This would be useful in facilitating the communication between Deaf and hearing persons. The goal is to go beyond the recognition of isolated signs or continuous signs in short sentences based on video, without the use of special equipment such as data gloves or magnetic markers. The focus was on the design of scalable formalisms for representation, model learning, and matching methods that are robust to image segmentation errors. |
|
Vision Problems in Automated Recognition of Sign Language (2003-current) |
|
Research Projects |
|
Cross-modal effects of enhanced perception exist in humans. There is also emerging experimental evidence about the neural basis of cross-modal perceptual organization. From an (Gibsonian) ecological point of view, it is reasonable to postulate that multi-sensory stimulation that co-varies in place and time originates from a single event or object. All of these point to the importance of cross-modal effects in the perception of the world by natural systems, which suggests that such principles and processes could be important in artificial systems. Of course, the representations and implementations of these ideas need not be constrained by those in the natural systems. The goal of the proposed research is to suggest representations, find computational models, and hypothesize semantic-level interpretations of audio-video events, by exploiting the perceptual organization of sounds and images, constrained by prior knowledge, as captured the ontology of the application domain. Perceptual organization of events will be used to bridge the semantic gap between the features extracted from the raw signal and the high-level, domain-dependent, semantics. |
|
Perceptual Organization of Image Features and Audio-Video Events |
|
The ability of being to identify humans from a distance in a passive manner has obvious applications in surveillance and threat assessment. However, there are other possible innovative uses, such as in smart rooms, designing environmentally aware electronic devices, and next generation computer games. In this general context, we are (i) researching modalities to recognize persons from a distance using image and video data, and (ii) looking into privacy and security related issues. In particular, we have found a novel way to reconstruct biometric templates from scores. This exposes a serious vulnerability in biometric systems |
|
Biometrics: Far, Outdoors, Security, and Privacy Issues (2001—current) |
|
The goal of this work was to develop a non-invasive imaging device based on regular 2-dimensional color images and 3-dimensional (3D) range images that can be used to collect data about the physical characteristics of human skin in terms of its color, texture, and elasticity. The applications for such measures are in domains where objective evaluation of skin condition is critical, such as in evaluating burn scars and in diagnosing skin melanoma. We used physics based computer vision methods to correct for image distortions due to incident illumination, location of light source, shading due to shape changes, and camera response non-linearities. We formulated theory to estimate relative elasticity using inverse-FEM methods. |
|
Human Skin Color and Elasticity from Images (1994-2002) |
|
Performance evaluation of computer vision algorithms is necessary not only to determine their limitations but also to facilitate the understanding of their underlying structure, eventually suggesting ways for their improvement. Empirical performance evaluation techniques allow us to tackle the problem of analysis of complex algorithms on real images. Among the major issues in empirical performance evaluation are (i) choice of the image set and ground truth, (ii) choice of the evaluation measures, (iii) strategy for selecting the parameters of the vision algorithm, and (iv) a thorough statistical analysis of the performance. |
|
Issues in Empirical Evaluation of Vision Algorithms (1995-2000) |
|
We proposed a new method of evaluating the effectiveness of low level computer vision algorithms such as edge detection and region segmentation. The general idea was to determine whether the algorithm produces information that is useful for human object recognition. We ran the algorithms on representative natural stimuli (e.g., photographs of objects in scenes), to produce images such as edge maps. Then we presented the resultant images to humans for recognition, and measured human performance. The level of performance was taken as an indication of the effectiveness of the algorithm: Algorithms that produced higher performance was viewed as providing more of the information that is useful in recognition. |
|
How Well Do Feature Extractors Support Object Recognition?(1994-1996) |
|
Field coupled computing is a radically different paradigm where electrical (QCA), magnetic (nano-magnets) or spin coupling among nano-devices are utilized for computation. We are exploring how traditional logic-based computing can be accomplished using such coupled devices. In addition, we are looking at an unconventional front in computing, which we call magnetic field-based computing (MFC), that harnesses the energy minimization aspects of a collection of nanomagnets to directly solve quadratic energy minimization problems, such as those arising in computationally intensive computer vision tasks. |
|
Computing with Nanodevices and Computer Vision |
|
Collaborator: Sanjukta Bhanja |
