Manuel Günther (here in the USA called Manuel Gunther; and sometimes seen as Manuel Guenther)
Resume: Curriculum Vitæ

Vision and Security Technology Lab (VaST)
1420 Austin Bluffs Parkway
P.O. Box 7150
Colorado Springs, CO 80933-7150

Current Occupation

Currently, I am working as a research associate at the Vision and Security Technology (VaST) lab in the University of Colorado Colorado Springs (UCCS) under the supervision of Prof. Terrance E. Boult. Here, I am part of the IARPA JANUS team investigating algorithms for unconstrained face recognition, including recognizing frontal and profile faces in unconstrained images and videos. The basic idea of my current research is to automatically estimate facial attributes (such as: big nose, bushy eye-brows, gender, ...) from facial images, and use these attributes as kind of a soft biometric to identify the person in the image.

As almost every researcher does nowadays, I am working with deep convolutional neural networks, and I have implemented several extensions to the Caffe framework, for example, a data layer to perform random rotation, scaling, shifting and blurring of training images in order to improve the stability of the network with respect to these transformations, which has been demonstrated in the AFAFCT paper at IJCB 2017, see below. Another such layer is a balancing layer, which balanced biased data distributions in multi-objective optimization problems, which we have published in the MOON paper at ECCV 2015, see below.

Of course, I am also actively participating in many other topics of the VaST lab, which are: adversarial images, improved network training, adversarial stability, and alike. For example, together with some colleagues, we were running the Unconstrained Face Detection and Open-Set Face Recognition Challenge, the results of which we present at IJCB 2017. Finally, I am still a developer of Bob (see below), though my activities there are a little limited lately. Anyways, I found the time to present the biometrics framework in a hands-on tutorial on 2D face recognition, held in the Biometrics Summer School 2016 in Kuala Lumpur, Malaysia.

Short Biography

In my early teenager times, I attended a high-school specialized for mathematics and natural science. As the logical consequence, I studied computer science with mathematics as minor subject at the Technical University of Ilmenau, Germany, starting in 1999. In my diploma thesis that I wrote in 2004, I also had my first contact with the face recognition topic, which I am following for more than 10 years now.

I wrote my PhD thesis about face detection, recognition, classification, and visualization in the Institute for Neural Computation at the Ruhr-University in Bochum, Germany, under the supervision of Dr. Rolf Würtz. There, I developed a simple statistical model that needs only few training data and can be trained and applied fast. Based on Gabor wavelet responses, I used my model to detect faces in images, recognized the person in the image, and classified, e.g., illumination conditions and facial expressions. I had implemented my model into the Pattern Recognition And Graph Matching Algorithms library, which unfortunately never was published as open source. Furthermore, I established a system to classify genetic syndromes based on facial images and implemented a graphical user interface called FIDA (Facial Image Diagnostic Aid). At the end of my time in Bochum, I stepped on novel Gabor-wavelet-based features that turned out to be very successful for face recognition and facial landmark localization under uncontrolled illumination conditions.

My next career step was leading me to the Idiap Research Institute, located in a small town Martigny in the middle of the Alps in Switzerland, where I joined the Biometrics group under the supervision of Dr. Sébastien Marcel. There, I was part of the development team of Bob, a free signal processing and machine learning toolbox for researchers, which is particularly designed to run biometric recognition experiments. The interface of Bob is written in Python, while the computationally intensive parts are implemented in C++ and bound to Python using the Python C-API. As the first step, I ported the Gabor wavelet based algorithms from my PhD thesis into Bob, resulting in the package bob.ip.gabor.
As another part of Bob, I developed the packages, an open source tool for the fair comparison of face recognition algorithms. This tool allows to run face (and other biometric) recognition experiments choosing from a variety of preprocessing algorithms, different features to extract from the images, and a pool of face recognition algorithms, and you can run the algorithm on one of the many available image and video databases. To assure comparability of the results, it is assured that default evaluation protocols are employed. The packages are designed in a way that makes it easy to implement your own preprocessor, feature, recognition algorithm or database, and you can simply plug in your code and combine it with the existing algorithms, and run an experiment to see whether your code works better than the existing one.
During my stay at Idiap, I became responsible for chairing the competition on face recognition in mobile environment using the MOBIO database. In this competition we found algorithms that can be used in real-world applications, i.e., when the image capture conditions cannot be controlled. The baseline script, which can be taken as an example, can be downloaded here. If you want to see, how my Gabor features (see above) performed in the competition, you might want to download the source code. The competition was held synchronized with the competition on speaker recognition. At the end of this competition, we fused the results of the best face and the best speaker recognition algorithms to build an integrated verification system. Details about the results of the competition and the fusion can be found in the ICB 2013 and BTFS 2013 papers listed below.

Professional Experience

The biggest field of research experience is on automatic face recognition. During my PhD time I became an expert on Gabor wavelet based face detection and recognition. I invented several methods to improve face detection, facial feature localization, and face recognition. I also successfully used Gabor wavelet responses for classification, inventing a statistical model that was able to deal with small amounts of training data and that handled Gabor wavelet responses in a proper way.
At Idiap, I have worked with a broader spectrum of face recognition algorithms. I have designed software and experiments to compare face recognition algorithms under different aspects such as facial expression, illumination, partial occlusion and non-frontal pose.
As I was giving a students programming exercise on Artificial Neural Networks, I am also familiar with that topic. But I am not convinced that Neural Networks (as they are currently used) are the best choice to solve problems.
Up till now, I was implementing all my experiments in C++ and Python, and I have some experience with Java programming, the Java Native Interface (JNI), the Python C-API, and also with using CMake, distributing mixed and inter-dependent C++/Python packages on the Python Package Index (PyPI), writing publications in LaTeX, and writing my personal web page in HTML. Of course, I also can work with Matlab and C, but I'd prefer to avoid these two languages as far as possible.