The problem of estimating dynamic information from static handwritten scripts is a joint effort between the following groups:
Groups |
---|
The Applied Mathematics group of the Department of Mathematical Sciences at the University of Stellenbosch in South Africa. |
The Digital Signal Processing (DSP) group of the Department of Electronic Engineering at the University of Stellenbosch in South Africa. |
Members of this research project include:
Members |
---|
Emli-Mari Nel received
the degree of Electrical and Electronic Engineering from the
University of Stellenbosch, South Africa in 2001. She received her
PhD in Electronic Engineering at the
University of Stellenbosch in 2005, under the supervision of Johan du Preez and Ben
Herbst. She has been working at Oxford Metrics Group as a researcher
in the Vicon Research Unit since October 2006. |
Johan A. du Preez joined the Department
of Electrical and Electronic
Engineering at the University of Stellenbosch in 1989 after four
years in the telecommunications sector. He received his PhD in
Electronic Engineering from the University of Stellenbosch in
1998. He is active in the broader fields of signal processing and
pattern recognition. His research focusses on developing advanced
statistical pattern recognition techniques such as higher-order
hidden Markov modeling and applying them to speech, on-line
handwriting and image processing problems. |
B. M. Herbst received his PhD in Applied Mathematics from the
University of the Free State, Bloemfontein, South Africa in 1982. He
visited the University of Dundee as a postdoctoral student during
1983 and was a visiting Associate Professor at Clarkson University,
Potsdam, NY (from 1988 to 1989) and the University of Colorado,
Boulder (from 1989 to 1990 and again during 1994). He joined the
Department of Applied Mathematics at the University of Stellenbosch
in 1998. His current research interests include scientific
computation, pattern recognition, and computer vision. |
[1] E. Nel, J. A. Du Preez, B. M. Herbst, Estimating the Pen Trajectories of Static Signatures using Hidden Markov models, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 27, 2005, pp. 1733-1746. [PDF]
[2] E. Nel, J. A. Du Preez, B. M. Herbst, Estimating the Pen Trajectories of Multi-Path Static Scripts using Hidden Markov Models, in the Proceedings of the International Conference on Document Analysis and Recognition, 2005, pp. 41-47. [PDF]
[4] J. A. Du Preez, Efficient Training of High-Order Hidden Markov Models using First-Order Representations, Computer Speech and Language, January 1998, Vol. 12, No. 1, pp. 23-39. [PDF]
[5] J. A. Du Preez, Efficient High-Order Hidden Markov Modelling, University of Stellenbosch, 1997. [PDF]
Our work on this research topic has been presented by Emli-Mari Nel at the following institutions: the International Conference on Document Analysis and Recognition in Seoul, South Korea, 2005 (ICDAR2005), the University of Colorado in Boulder, Colorado, 2005, the annual conference of the South African Society for Numerical and Applied Mathematics (SANUM), 2003, 2005, and 2006, and at Microsoft Research in Redmond, 2005.
A general presentation containing most of the material of the presentations above can be downloaded [here]. See the enclosed README.txt file for information on this presentation. (Some animations are embedded in slides 43 and 49 which can be uploaded by clicking on the appropriate buttons/pictures.)
US_SIGBASE is a database that contains both the on-line and off-line versions of 51 signatures. Additionally, it contains approximately 14 other dynamic representatives of the available off-line signatures. See the enclosed "README.txt" file for more information, after downloading US_SIGBASE.tar.gz. See [1], [2] and [3] for publications that used the US_SIGBASE to generate their results. Emli-Mari Nel, Johan du Preez, Ben Herbst developed the algorithms and software to record US_SIGBASE. For any enquiries, feel free to contact us. US_SIGBASE is especially useful to estimate dynamic information from static handwritten scripts. Animations were created to visualize results from our pen-trajectory estimations algorithm.
This section contains examples of our pen-trajectory estimation algorithm. Our algorithm estimated the path followed by the pen, i.e., the pen trajectory, to create each of the shown images.
For each image the estimated pen trajectory can be viewed as an animation. As mentioned above, the dynamic counterpart is available for each static image in the US_SIGBASE. These dynamic counterparts were NOT used to estimate our trajectories but to quantitatively evaluate the efficacy of our estimated pen trajectories. The dynamic counterparts (equivalents) of each static image were used to estimate the ground truths for each image which can also be viewed by clicking on any of the signatures below. Subsequently, our evaluation protocol were used to establish a pointwise comparison between the estimated pen trajectories and their ground truths.
Our evaluation protocol enables one to identify the local errors in an estimated pen trajectories. These errors are subdivided into substitution, insertion and deletion errors. These errors can also be seen by clicking on any of the signatures.
See [1], [2] and [3] for publications on this topic. Emli-Mari Nel, Johan du Preez, Ben Herbst developed the algorithms and software for the shown examples.
We have developed an algorithm that computes that optimal mapping between an estimated trajectory and its ground-truth. Both sequences should be presented as a sequence of edges. The estimated trajectories from our recovery algorithm were compared to their ground-truth trajectories (from US_SIGBASE ). Click on any of the above images to view animations of the local errors that occurred in the esimated trajectories. We will gladly provide results from this evaluation algorithm to other researchers on request .