Thor List

Comparison of target detection algorithms using adaptive background models

This article compares the performance of target detectors based on adaptive background differencing on public benchmark data. Five state of the art methods are described. The performance is evaluated using state of the art measures with respect to ground truth. The original points are the comparison to hand labelled ground truth and the evaluation on a large database. The simpler methods LOTS and SGM are more appropriate to the particular task as MGM using a more complex background model.

Performance evaluating the evaluator

When evaluating the performance of a computer-based visual tracking system one often wishes to compare results with a standard human observer. It is a natural assumption that humans fully understand the relatively simple scenes we subject our computers to and because of this, two human observers would draw the same conclusions about object positions, tracks, size and even simple behaviour patterns. But is that actually the case? This paper provides a baseline for how computer-based tracking results can be compared to a standard human observer.

Efficient Hidden Semi-Markov Model Inference for Structured Video Sequences

The semantic interpretation of video sequences by computer is often formulated as probabilistically relating lower-level features to higher-level states, constrained by a transition graph. Using hidden Markov models inference is efficient but time-in-state data cannot be included, whereas using hidden semi-Markov models we can model duration but have inefficient inference. We present a new efficient O(T) algorithm for inference in certain HSMMs and show experimental results on video sequence interpretation in television footage to demonstrate that explicitly modelling time-in-state improves interpretation performance

CVML - an XML-based computer vision markup language

We propose an XML-based computer vision markup language for use in cognitive vision, to enable separate research groups to collaborate with each other as well as making their research results more available to other areas of science and industry, without having to reveal any proprietary ideas, algorithms or even software. The computer vision markup language can communicate any type and amount of information, making unavailable functionality accessible to anyone. We introduce the language and describe how we have implemented it in a very large cognitive vision project. We provide a free open source library for working with this language, which can easily be implemented into existing code providing seamless network communication abilities and multi-platform support. Finally, we describe the future of CVML and how it might evolve to include other areas of research.

Artificial intelligence in computer graphics: a constructionist approach

Introduction The ultimate embodiment of artificial intelligence (AI) has traditionally been seen as physical robots. However, hardware is expensive to construct and difficult to modify. The connection between graphics and AI is becoming increasingly strong, and, on the one hand, it is now clear that AI research can benefit tremendously from embodiment in virtual worlds [1, 4, 5]. On the other hand, computer games, a highly visible area of computer graphics development, could benefit from the use of more advanced AI [11]. But a synergistic marriage of the two fields has yet to occur. A substantial impediment to introducing more intelligent characters into games is the lack of practitioners who understand both the realms of graphics and of AI, and who can drive their integration. Another is the diversity in programming languages and environments that people from both camps use. We are working to address these problems using a multi-prong approach. Here we present two of these, which are interrelated. First, we present a new design methodology aimed at making the construction of AI systems easier for novices and experts alike. Second, we present network-based software designed to take advantage of this methodology, allowing general-purpose systems-level integration of AI programs with other systems, including graphics. The software is available free of charge to researchers. An underlying assumption for this work is the theory that the mind can be modeled through the adequate combination of interacting, functional machines, or modules. In his seminal book, Unified Theories of Cognition, the late Allen Newell (1990) [16] called on researchers to start working on unified theories of mind, rather than continuing working with micro-theories about isolated components. We agree strongly with Newell in his quest for integration, and we do not interpret his call to mean a corollary to the search for the unified field theory in physics. Rather, we subscribe to the implications of Minsky’s (1986) Society-of-Mind theory [8] that the mind is a multitude of diverse, interacting components. Our methodology and software foundation directly supports the construction of such systems. The motivations for this work are numerous. Large systems such as virtual worlds with simulated inhabitants cannot be built from scratch without bringing together a large team of experts in each field that such a system naturally encompasses. Our methodology aims to help coordinate the effort. There is also a lack of incremental accumulation of knowledge in AI and related computer graphics. By supporting reuse of prior work we enable the building of increasingly powerful systems, as core system elements do not need to be built from scratch.

A plug-and-play architecture for cognitive video stream analysis

This paper presents an architecture for cognitive analysis of streaming video, in which a new module can easily be plugged in, to add to or even compete with existing functionality. This allows the implementers to focus on the key scientific issues instead of struggling with the details of the implementation. The architecture is distributed and runs independently of the underlying computer architecture and can run transparently across one or many different operating systems in a larger distributed system. This architecture focuses on several key computer vision issues, such as multi-level global and local control, automatic dataflow based on auto-descriptive self-regulating independent modules that come together to form a whole based on the characteristics of the individual and the needs of the system rather than a static flow diagram.

An intelligent and task-independent controller for video sequence analysis

This paper describes a task-independent controller that allows for an easy implementation of vision systems for processing video sequences. The controller does not have a fixed dataflow or any fixed steps. The dataflow is constructed by the modules by describing themselves for the controller. During operation the modules and their parameters are selected using an independent decision module. This makes the system flexible and allows comparison of different learning techniques and decision strategies. The controller is being used by the CAVIAR system and its current decision module is a rule-based system written in Clips.