Jacobus J Barnard

Abstract:

There is a growing trend in machine color constancy research to use only image chromaticity information, ignoring the magnitude of the image pixels. This is natural because the main purpose is often to estimate only the chromaticity of the illuminant. However, the magnitudes of the image pixels also carry information about the chromaticity of the illuminant. One such source of information is through image specularities. As is well known in the computational color constancy field, specularities from inhomogeneous materials (such as plastics and painted surfaces) can be used for color constancy. This assumes that the image contains specularities, that they can be identified, and that they do not saturate the camera sensors. These provisos make it important that color constancy algorithms which make use of specularities also perform well when the they are absent. A further problem with using specularities is that the key assumption, namely that the specular component is the color of the illuminant, does not hold in the case of colored metals. In this paper we investigate a number of color constancy algorithms in the context of specular and non-specular reflection. We then propose extensions to several variants of Forsyth's CRULE algorithm^1-4 which make use of specularities if they exist, but do not rely on their presence. In addition, our approach is easily extended to include colored metals, and is the first color constancy algorithm to deal with such surfaces. Finally, our method provides an estimate of the overall brightness, which chromaticity-based methods cannot do, and other RGB based algorithms do poorly when specularities are present.

Abstract:

We consider object recognition as the process of attaching meaningful labels to specific regions of an image, and propose a model that learns spatial relationships between objects. Given a set of images and their associated text (e.g. keywords, captions, descriptions), the objective is to segment an image, in either a crude or sophisticated fashion, then to find the proper associations between words and regions. Previous models are limited by the scope of the representation. In particular, they fail to exploit spatial context in the images and words. We develop a more expressive model that takes this into account. We formulate a spatially consistent probabilistic mapping between continuous image feature vectors and the supplied word tokens. By learning both word-to-region associations and object relations, the proposed model augments scene segmentations due to smoothing implicit in spatial consistency. Context introduces cycles to the undirected graph, so we cannot rely on a straightforward implementation of the EM algorithm for estimating the model parameters and densities of the unknown alignment variables. Instead, we develop an approximate EM algorithm that uses loopy belief propagation in the inference step and iterative scaling on the pseudo-likelihood approximation in the parameter update step. The experiments indicate that our approximate inference and learning algorithm converges to good local solutions. Experiments on a diverse array of images show that spatial context considerably improves the accuracy of object recognition. Most significantly, spatial context combined with a nonlinear discrete object representation allows our models to cope well with over-segmented scenes. © Springer-Verlag 2004.

Abstract:

Recreation simulation modeling, when combined with intelligent monitoring, is becoming a valuable tool for natural resource managers. The goal of recreation simulation is to accurately model recreational use, both current and future. Models are applied to gain a thorough understanding of the characteristics of recreation. Indicator variables such as visitor experience, carrying capacity and impact on resources can be computed. If the model is valid it can be used to predict future use as well as to investigate the effect of new scenarios and management decisions. Recent research has focused on agent-based modeling techniques. Recreators are represented by autonomous, intelligent agents that travel across the landscape. A central issue is the model used for agent travel decisions. Current techniques range from replicating trips exactly to making local, intersection level decisions based on probability. But little attention has been paid to justifying these models. In this work we examine a range of probabilistic models. The models differ in the length of the Markov chain used to compute agent decisions. The length of the chain ranges from zero (local decisions only) to infinity (exact trip replication). We test the length of the chain on held out data for validation. We show that the choice of model strongly influences the validity and results of the simulation. To test these models we present a framework for automatically constructing agent-based models from an input set of GPS tracklogs. The GPS tracklogs are collected by volunteers as they recreate in natural areas. Traditionally, data on where recreators travel is collected in the form of trip diaries, filled out on paper by visitors or by interview. Other demographic and attitudinal data is also collected along with the actual route traveled. Although the additional information is valuable, the data must be collected and entered by hand. Paper diaries also place a significant time burden on visitors, reducing the compliance rate as well as skewing the results (ensuring only visitors with excess time participate). Using GPS devices to record visitor trips helps alleviate these problems. The framework for processing GPS trips and automatically building a model presented in this work significantly reduces the time required to build a model, lowers the cost and widens the applicability of recreation simulation modeling to new areas. GPS devices automatically record their data, requiring only that visitors turn the unit on and carry it with a marginal view of the sky. GPS use is also becoming more widespread among recreators. As more recreators use GPS to record their trips, data useful to modeling is becoming increasingly available. The steps in GPS driven model generation are as follows. First, the set of GPS tracklogs is combined to form the underlying travel network along which agents will travel. Each GPS tracklog is then traced along the network in order to determine what choices were made as the recreator traveled across the network. This produces a list of trip itineraries. Model parameters (probability tables) can then be computed from the trips. The length of the Markov chain used in the probability tables is a parameter to the model. The optimal value is found by testing the likelihood of heldout data for different chain lengths. This step is done automatically. Once the optimal length of the chain is chosen the model is complete and agent-based simulation can proceed. The entire framework for automatically producing GPS driven agent-based models is implemented in our TopoFusion GPS mapping software. We present results from two collections of GPS tracklogs from different trail systems. The first is from Tucson Mountain Park and is the result of a volunteer collection effort by the authors. A trails master plan is underway at the park, with input from our model. The second is a collection of tracks from mountain bike rides in the Finger Rock Wash area, collected by the author. Testing by held-out data on both GPS datasets indicates that current modeling methods are insufficient to model recreator travel decisions. The middle ground (neither exact replication nor local decisions) consistently performs better.

Abstract:

Multi-touch tablets can offer a large, collaborative space where several users can work on a task at the same time. However, the lack of privacy in these situations makes standard password-based authentication easily compromised. This work presents a new gesture-based authentication system based on users' unique signature of touch motion when drawing a combination of one-stroke gestures following two different policies, one fixed for all users and the other selected by a model of control to maximize the expected long-term information gain. The system is able to achieve high user recognition accuracy with relatively few gestures, demonstrating that human touch patterns have a distinctive "signature" that can be used as a powerful biometric measure for user recognition and personalization. © 2013 Springer-Verlag Berlin Heidelberg.

Abstract:

We present a novel method for detecting and quantifying 3D structure in stacks of microscopic images captured at incremental focal lengths. We express the image data as stochastically generated by an underlying model for biological specimen and the effects of the imaging system. The method simultaneously fits a model for proposed structure and the imaging system's parameters, which include a model of the point spread function. We demonstrate our approach by detecting spores in image stacks of Alternaria, a microscopic genus of fungus. The spores are modeled as opaque ellipsoids and fit to the data using statistical inference. Since the number of spores in the data is not known, model selection is incorporated into the fitting process. Thus, we develop a reversible jump Markov chain Monte Carlo sampler to explore the parameter space. Our results show that simultaneous statistical inference of specimen and imaging models is useful for quantifying biological structures in 3D microscopic images. In addition, we show that inferring a model of the imaging system improves the overall fit of the specimen model to the data. © 2006 IEEE.