Vision Cohort 2019

inzva held a one-of-a-kind event, Vision Cohort, that allowed acclaimed academics and students working in the field of computer vision from various universities located in different cities to come together to have fruitful discussions about their works and the computer vision community in Turkey. Building bridges between Turkey’s two major cities, Ankara and Istanbul; inzva invited a total of ten academics and seventeen graduate-level students to share their ongoing projects, and their ideas for the future.

The weekend-long event was full of brainstorming, and ultimately proved to be a great opportunity to meet like-minded scholars, all working to contribute to the field with their own expertise.

While the academics talked about their most recent work and relevant research during the presentations in the mornings, the students held poster sessions to receive feedback from both the academics and their peers to further improve their work.

Our visitors not only had an intensive weekend that gave them insight into different aspects of their field; but also discovered Beykoz Kundura where inzva is located, and got to enjoy spending time with their families in the historical buildings, their surroundings and the beautiful view of the sea.

You can see the details of the presentations given by our invited academics in order of appearance below:

Assoc. Prof. Hazım Kemal Ekenel from Istanbul Technical University gave his talk on the recent work conducted in his computer vision laboratory at ITU, SiMiT Lab. Ekenel and his students are working on problems such as transfer learning, age and gender classification, 2D to 3D face modeling and various others. During his talk, Dr. Hazım emphasized the importance of considering dataset bias while studying a particular problem using the same dataset. Domain adaptation and context adaptation on deep learning was also discussed.

Prof. Pınar Duygulu Şahin from Hacettepe University Computer Vision Lab (HUCVL) talked about several recent academic studies of her and her students’. One recent work is image to image translation from actual images to illustrations with the style of particular illustrators. Şahin and colleagues used image style transfer techniques to achieve this. The same dataset gathered from children’s illustrated books is also used for classifying the illustrators. Şahin also talked about iterative data organizing and cleaning methods to improve video search results with text-based queries by leveraging weakly labeled data.

Prof. Gözde Ünal from ITU Computer Vision and Artificial Intelligence Laboratory (ITU Vision Lab) talked about the importance of mathematical modeling in computer vision. Specifically, Ünal presented their work on color-coding the nerve tracts from MRI images using mathematical modeling and geometric representations. Ünal underlined that it is not always smarter to overshadow mathematical and geometric modeling by hastily applying deep learning methods to solve problems. She emphasized that many classical approaches are still able to solve novel problems and should not be forgotten; and that we shouldn’t overlook their efficiency while dealing with more recent problems. 

Assoc. Prof. Aykut Erdem is from Hacettepe University Computer Vision Lab (HUCVL). His talk was based on reasoning through neural networks. He started by giving background information on the resurgence of neural networks. Then he talked about the RecipeQA dataset which was prepared by him and his students to study the reasoning abilities of deep neural networks. He presented a novel deep learning architecture named Procedural Reasoning Networks that use relational recurrent neural networks and bidirectional attention flow to understand and answer questions related to the cooking recipes.

Assoc. Prof. Erkut Erdem of Hacettepe University Computer Vision Lab (HUCVL) talked about their recent work on language based natural scene manipulation. Erdem and his colleagues used an architecture based on conditional generative adversarial networks (GANs) to perform scene manipulation. The GAN is conditioned to different attributes of scenes such as grass, mountains, trees, and the amount of sea in an image. They get excellent results on translating a scene image with its original attributes to a scene with different attributes (such as a scene in winter to a scene in summer) with convincingly good photorealism.

Assoc. Prof. Seniha Esen Yüksel Erdem is the director of the Pattern Recognition and Remote Sensing Laboratory (PARRSLAB) in the Department of Electrical and Electronics Engineering at Hacettepe University. Her talk covered her recent work on super-resolution for hyperspectral images. Hyperspectral images give information about the presence of different materials on a surface. Yüksel and her colleagues proved that the super-resolution problem can be solved as a quadratic optimization problem and a global optimum can be found.

Assoc. Prof. Çiğdem Gündüz Demir is from the Department of Computer Engineering at Bilkent University. Gündüz made a presentation about using deep learning methods for medical image analysis. She began her presentation by explaining why computer-aided methods and algorithms are needed in the medical image analysis domain. She presented examples of microscopic images and explained why the analysis (e.g., cell counting) is a difficult problem, both for the human eye and computers. She then explained three different studies on cell and tissue segmentation that her research group developed using deep learning.

Asst. Prof. Hamdi Dibeklioğlu is a faculty member in the Department of Computer Engineering at Bilkent University. Dibeklioğlu presented research conducted by him and his colleagues on the analysis of hereditary and behavioral characteristics of facial expressions. He presented a study about telling if a person is lying, only by analyzing the person’s facial expressions. He described how to analyze human behavior for assessing depression severity. He also showed the study of predicting how a couple’s child’s expressions will look like by analyzing the facial expressions of the mother and the father.

Dr. Erdem Yörük is the chief scientist and a partner of Vispera and co-advises graduate students and offers a graduate-level course in the Department of Electrical and Electronics Engineering at Boğaziçi University. Yörük talked about the problem of detecting and recognizing the brands and types of products on the supermarket shelves for stock keeping. He showed great accuracy rates on both detecting and recognizing using deep learning methods. Yörük also emphasized the fact that Bayesian methods can be employed for more efficiency.

We concluded the weekend with a feedback session in which we talked about the upcoming projects we envisioned for the computer vision community and the AI community in general; and invited seven students chosen by the Ankara-based invited academics to our AI Projects #3 to collaborate with and for them to visit us each month for the next three months.

inzva is supported by BEV Foundation, an education foundation for the digital native generation which aims to build communities that foster peer-learning and encourage mastery through one-to-one mentorship.


Subscribe to our newsletter here and stay tuned for more hacker-driven activities.