Erdem Yörük, Ph.D.
I am a researcher and an academic concentrated in the fields of statistical modeling, machine learning, computer vision and computational biology. I received my BS and MS degrees in Electrical and Electronic Engineering from Bogazici University, and my PhD degree in Applied Mathematics and Statistics from Johns Hopkins University under the supervision of Prof. Donald Geman with an official concentration in the Mathematical Bioinformatics track of JHU-Institute for Computational Medicine. I continued my postdoctoral studies at JHU-Center for Imaging Science (CIS) and JHU-Vision Lab in joint projects with UC-Berkeley and UC-Los Angeles.
Currently, I am a partner and senior research scientist at Vispera Information Technologies Co., an Istanbul based startup company specialized in computer vision solutions in the FMCG retail industry. I lead the research, design, and development of the company’s machine learning and computer vision technology aimed at parsing complex and crowded scenes of retail stores to the scrutiny of individual packaged products. The patent-pending approach I work on can efficiently recognize, localize and count instances of several hundreds of different product brands captured in unconstrained imagery, such as mobile camera images; and enables a huge leap in automation and cost of real-world tasks such as inventory tracking, stock-out detection, and compliance assessment of retail shelves. Apart from my industrial post, I am also an adjunct faculty at Bogazici University - Department of Electrical and Electronics Engineering, where I co-advise graduate students and offer a graduate level course on Probabilistic Graphical Models with theory and applications to computer vision, machine learning and computational biology.
My other research activities including my doctoral and postdoctoral work and their ongoing extensions can be summarized as follows:
Small Sample Learning: In my doctoral dissertation, I addressed a very challenging machine learning problem common to most disciplines, namely discovering generalizable dependency structures from very high dimensional but scarcely sampled data. I developed a novel family of latent variable forest models and laid out the corresponding learning algorithm dubbed as stepwise dependency pursuit, which achieved state-of-the-art performance in hand-written character recognition, statistical shape modeling and cancer-type classification.
Statistical Modeling of Protein Signaling Networks: My doctoral work also extended to a comprehensive statistical model that can elucidate disease mechanisms of protein signaling networks. My framework can capture the entire generative process from disease phenotypes to final expression readings with sub-models specific to population, patient, tissue, cell and measurement levels. As shown in a clinical breast cancer study, it can successfully predict protein/receptor aberrations and unknown states of potential drug targets in pathways, by just observing gene expression microarray data of the signaling targets.
Adaptive Scene Understanding: During my postdoctoral studies I worked on adaptive interpretation of complex and cluttered scenes involving multiple interacting objects. The framework implements evidence integration and selective attention of human vision by combining learned scene priors with a battery of off-the-shelf object detectors. In a novel inference algorithm, called Entropy Pursuit, object deetectors are executed sequentially to test different pose (scale, location, orientation) and category hypotheses that when verified would most reduce the posterior entropy about the so far revealed scene content. This work is still ongoing with extensions to different context priors and scene types including my industrial domain of retail scenes.
3D Pose Estimation from 2D: Another ongoing project I have started during my postdoc is on 3D object detection and 6 degrees-of-freedom continuous pose estimation from a single 2D visible light image. I devised a novel approach that uses simple wireframe models learned from 2D object blueprints and composed of 3D edge primitives, and implements a Branch-Bound search for the globally optimal pose by maximizing the match between model’s camera projection and a novel quantized form of HOG features of the input image. The method has surpassed state-of-the-art in 3D detection and continuous pose estimation of cars, with a multiple orders of magnitude higher efficiency and received Microsoft best paper award at ICCV-3dRR. Currently, this project is being extended to multiple object categories, probabilistic object models and learning of 3D object models from real and (un)annotated images.
- Statistical Modeling: Probabilistic Graphical Models, Bayesian Networks, Causal Inference, Latent Variable Models, Small Sample Learning, Monte Carlo Methods
- Computer Vision: 3D Vision, Information Theoretic Scene Interpretation, Generative Models for Images, Learning Scene and Image Priors, Medical Imaging, Biometrics
- Bioinformatics: Biomolecular Network Inference, Protein Signaling Networks, Biomarker Discovery, Oncology Biostatistics
- D. Rother, E. Yörük, S. Mahendran, R. Vidal, “Hypothesize and Bound: A Computational Focus of Attention Mechanism for Simultaneous 3D Shape Reconstruction, Pose Estimation and Classification from a Single 2D Image" International Journal of Computer Vision, 2014. under revision.
- E. Yörük, R. Vidal, “Efficient Object Localization and Pose Estimation with 3D Wireframe Models" International Conference on Computer Vision - 4th International IEEE Workshop on 3D Representation and Recognition (ICCV - 3dRR), 2013.
- E. Yörük, M. F. Ochs, D. Geman, L. Younes, “A Comprehensive Statistical Model for Cell Signaling” IEEE Transactions on Computational Biology and Bioinformatics, 8(3), 592-606, 2011.
- B. Acar, E. Yörük, “DT-MRI Connectivity and/or Tractography?: Two New Algorithms”, Tensors in Image Processing and Computer Vision, S. Aja-Fernández, R. de Luis García, D. Tao, & Li, X. (Eds.), Series: Advances in Pattern Recognition, Springer, 2009.
- H. Dutağacı, G. Fouquier, E. Yörük, B. Sankur, L. Likforman-Sulem, J. Darbon, “Hand Recognition,” Guide to Biometric Reference Systems and Performance Evaluation, D. Petrovska-Delacretaz, G. Chollet, B. Dorizzi (Eds.), pp. 89-124, Springer, 2009
- H. Dutağacı, B. Sankur, E. Yörük, “A comparative analysis of global hand appearance-based person recognition,” Journal of Electronic Imaging, 17(1), 011018/1- 011018/19, Jan-March, 2008.
- E. Yörük, H. Dutagaci, B. Sankur, “Hand Biometrics,” Image and Vision Computing, 24(5), 483-497, 2006.
- E. Yörük, E. Konukoglu, B. Sankur, J. Darbon, “Shape-Based Hand Recognition,” IEEE Transactions on Image Processing, 15(7), 1803-1815, 2006.
- E. Yörük, B. Acar, R. Bammer “A Physical Model for DT-MRI Based Connectivity Map Computation” Lecture Notes in Computer Science, Vol. 3749 / Part 1, pp. 213 220, Springer, 2005 (MICCAI 2005, Palm Springs, CA, USA)
- E. Yörük, B. Acar, “Structure Preserving Regularization of DT-MRI Vector Fields By Nonlinear Abisotropic Diffusion Filtering” Proceedings of EUSIPCO 2005, European Signal Processing Conference, September 4-8, 2005, Antalya, Turkey
- E. Yörük, H. Dutagaci, B. Sankur “Hand Based Biometry” Proceedings of SPIE 2005, January 16-20, 2005, San Jose, CA, USA
- E. Yörük, E. Konukoglu, B. Sankur, J. Darbon, “Person Authentication Based On Hand Shape”, Proceedings of EUSIPCO 2004, 12th European Signal Processing Conference, September 6-10, 2004, Vienna/Austria
- E. Yörük, C. B. Akgül, “Color Image Segmentation using PDE-based Regularization and Watershed Transformation”, Proceedings of IEEE-SIU 2004, IEEE 12th Signal Processing and Telecommunication Applications Conference, April 28-30, 2004, Kusadasi/Turkey
- E. Yörük, E. Konukoglu, B. Sankur, J. Darbon, “Shape Based Hand Recognition”, Proceedings of IEEE-SIU 2004, IEEE 12th Signal Processing and Telecommunication Applications Conference, April 28-30, 2004 Kusadasi/Turkey (ranked first in student paper contest)
- JHU Center for Imaging Science:
- JHU Institute For Computational Medicine:
- JHU The Vision, Dynamics and Learning Lab:
- JHU Applied Mathematics and Statistics:
- Bogazici University Electrical Engineering: