Abstract

Writer independent handwriting recognition systems are limited in their accuracy, primarily due the large variations in writing styles of most characters. Samples from a single character class can be thought of as emanating from multiple sources, corresponding to each writing style. This also makes the inter-class boundaries, complex and disconnected in the feature space. Multiple kernel methods have emerged as a potential framework to model such decision boundaries effectively, which can be coupled with maximal margin learning algorithms. We show that formulating the problem in the above framework improves the recognition accuracy. We also propose a mechanism to adapt the resulting classifier by modifying the weights of the support vectors as well as that of the individual kernels. Experimental results are presented on a data set of 16,000 alphabets collected from 470 writers using a digitizing tablet

Abstract

Widespread use of surveillance cameras in offices and other business establishments, pose a significant threat to the privacy of the employees and visitors. The challenge of introducing privacy and security in such a practical surveillance system has been stifled by the enormous computational and communication overhead required by the solutions. In this paper, we propose an efficient framework to carry out privacy preserving surveillance. We split each frame into a set of random images. Each image by itself does not convey any meaningful information about the original frame, while collectively, they retain all the information. Our solution is derived from a secret sharing scheme based on the Chinese Remainder Theorem, suitably adapted to image data. Our method enables distributed secure processing and storage, while retaining the ability to reconstruct the original data in case of a legal requirement. The system installed in an office like environment can effectively detect and track people, or solve similar surveillance tasks. Our proposed paradigm is highly efficient compared to Secure Multiparty Computation, making privacy preserving surveillance, practical.

Abstract

Fourier domain methods have had a long association with geometric vision. In this paper, we introduce Fourier domain methods into the field of visual servoing for the first time. We show how different properties of Fourier transforms may be used to address specific issues in traditional visual servoing methods, giving rise to algorithms that are more flexible. Specifically, we demonstrate how Fourier analysis may be used to obtain straight camera paths in the Cartesian space, do path following and correspondenceless visual servoing. Most importantly, by introducing Fourier techniques, we set a framework into which robust Fourier based geometry processing algorithms may be incorporated to address the various issues in servoing.

Abstract

Super-resolution reconstruction algorithms assume the availability of exact registration and blur parameters. Inaccurate estimation of these parameters adversely affects the quality of the reconstructed image. However, traditional approaches for image registration are either sensitive to image degradations such as variations in blur, illumination and noise, or are limited in the class of image transformations that can be estimated. We propose an accurate registration algorithm that uses the local phase information, which is robust to the above degradations. We derive the theoretical error rate of the estimates in presence of non-ideal band-pass behavior of the filter and show that the error converges to zero over iterations. We also show the invariance of local phase to a class of blur kernels. Experimental results on images taken under varying conditions clearly demonstrates the robustness of our approach.

Abstract

We propose a novel approach to 3D hand geometry based person authentication using projected light patterns. Instead of explicitly computing a depth map of the palm for recognition, we capture the depth information in the deformations of a projected texture pattern, and use it directly for recognition. The deformed pattern is characterized using local texture measures, which can encode the certain depth characteristics of the palm. An authentication system built using the proposed technique achieves an equal error rate of 0.84% on a dataset of 1341 samples collected from 149 users, as opposed to 4.03% using traditional 2D features on an identical dataset. The approach is robust as well as computationally efficient and could be applied to other 3D object recognition problems as well.

Abstract

Text-dependent writer verification systems are preferred over text-independent systems due to the accuracy they achieve with small amount of data. However, text-dependent systems are prone to forgery. This paper proposes a novel boosting based framework for writerspecific text generation to increase the accuracy and a method of text variation to make the system robust to forgery. The approach is able to achieve error rates of 5% with just 6 words as compared to random(11%) or most discriminative(22%) primitive selection methods on a dataset containing 20 writers. Boosting based text selection also provides the flexibility to incorporate text variation across multiple authentications, which in turn makes the system robust to forgery

Abstract

Algorithms for classification of 3D objects either recover the depth information lost during imaging using multiple images, structured lighting, image cues, etc. or work directly the images for classification. While the latter class of algorithms are more efficient and robust in comparison, they are less accurate due to the lack of depth information. We propose the use of structured lighting patterns projected on the object, which gets deformed according to the shape of the object. Since our goal is object classification and not shape recovery, we characterize the deformations using simple texture measures, thus avoiding the error prone and computationally expensive step of depth recovery. Moreover, since the deformations encode depth variations of the object, the 3D shape information is implicitly used for classification. We show that the information thus derived can significantly improve the accuracy of object classification algorithms, and derive the theoretical limits on height variations that can be captured by a particular projector-camera setup. A 3D texture classification algorithm derived from the proposed approach achieves a ten-fold reduction in error rate on a dataset of 30 classes, when compared to state-of-the-art image based approaches. We also demonstrate the effectiveness of the approach for a hand geometry based authentication system, which achieves a four-fold reduction in the equal error rate on a dataset containing 149 users.

Abstract

Projector-Camera systems are extensively used for various applications in computer vision, immersive environments, visual servoing, etc. Due to gaps between neighboring pixels on the projector’s image plane and variations in scene depth, the mage projected onto a scene shows pixelation and blurring artifacts. In certain vision and graphics applications, it is desirable that a high quality composition of the scene and the projected image, excluding the artifacts, is captured, while retaining the scene characteristics. Localization of projected pixels can also help in dense estimation of scene shape. In this paper, we address the problem of localizing each of the projected pixels from a captured scene and restoring the captured image so that the pixelation and blurring artifacts of the projector are removed. We improve the quality of the captured image further by virtualizing a high resolution projector. i.e., we modify the captured image as if the scene were illuminated with a high-resolution projector. We propose robust solutions to these problems and demonstrate their effectiveness on scenes of different complexities.

Abstract

Constructing a high-resolution (HR) image from low resolution (LR) image(s) has been a very active research topic recently with focus shifting from multi-frames to learning based single-frame super-resolution (SR). Multi-frame R algorithms attempt the exact reconstruction of reality, but are limited to small magnification factors. Learning based SR algorithms learn the correspondences between LR and HR patches. Accurate replacements or revealing the exact underlying information is not guaranteed in many scenarios. In this paper we propose an alternate solution. We propose to capture images at right zoom such that it has just sufficient amount of information so that further resolution enhancements can be easily achieved using any off the shelf single-frame SR algorithm. This is true under the assumption that such a zoom factor is not very high, which is true for most man-made structures. The low-resolution image is divided into small patches and ideal resolution is predicted for every patch. The contextual information is incorporated using a Markov Random Field based prior. Training data is generated from high-quality images and can use any singleframe SR algorithm. Several constraints are proposed to minimize the extent of zoom-in. We validate the proposed approach on synthetic data and real world images to show the robustness

Abstract

Recognition of object categories from their images is extremely challenging due to the large intra-class variations, and variations in pose, illumination and scale, in addition to lack of depth information of the object. Recovering the depth information from multiple images or from image cues such as variations in illumination or focus, is both computationally intensive and error prone. In contrast, the appearance based approaches are more robust and computationally efficient. However, they lack the potential accuracy of 3D feature based approaches due to the lack of shape information. We propose the use of structured lighting patterns projected on the object, which gets deformed according to the shape of the object for recognition. Since our goal is object classification and not shape recovery, we characterize the deformations using simple texture measures, thus avoiding depth recovery step. Moreover, the shape information present in the deformations is implicitly used for classification. We show that the information thus derived can significantly improve the accuracy of object category recognition from arbitrary-pose images.