Features extracted from speech are widely used for problems such as biometric speaker identification, but the use of speech data raises concerns about privacy. We propose a method for speech and posture classification using only breath data. The acoustical information was extracted from breath instances using the Hilbert-Huang transform and fed into our CNN-RNN network for classification. We also created our publicly available dataset, BreathBase, which contains more than 5000 breath instances of 20 participants in 5 different postures with 4 different microphones. Using this data, 85% speaker classification and 98% posture classification accuracy is obtained.

Date: 27.11.2019 / 15.00 Place: A-212

There is a high demand for realistic computer aided imagery by many applicatiion areas such as games and movies. Due to the complicated characteristics of certain natural phenomena such as fire, smoke or mist, it is difficult to realistically mimic these effects. There are various approximation methods to visually synthesize lifelike 3D artifacts. The use of flow fields to guide the motion of particles creates a random but natural-looking effect. The aim of this study is to use flow fields to generate realistic 3D visual effects.

Date: 06.12.2019 / 13:00 Place: B-116

In this study, firstly, the effect of kinemorphs on F-formation shapes are investigated by focusing on the differences between the virtual environment and the real life settings. Secondly, the role of one-to-one F-formation shapes on joining a dyadic interaction as a third interactant is also studied.

Date: 12.09.2019 / 15:00 Place: A-212

Our project brings together the concept of sound generation from free text input. Because a good story text can describe an environment quite well, both in terms of visual and aural aspects, we decided to use this feature to create virtual soundscapes that will both be generated procedurally and have the depth, complexity and control of a story text. In this thesis we describe the mechanics of analyzing free text user input queries and a way of converting this analysis into procedural soundscape generation.

Date: 05.09.2019 / 10:00 Place: A-108

In this study we implement a tool that compiles the combinatory categorial grammar rules to extract type-raising rules out of verb categories’ arguments and use them to test it out on natural language corpora like Eve.

Date: 05.09.2019 / 15:00 Place: S-06 (Conference Hall - 2)

Image based CAPTCHA systems are being used to validate if the candidate user is a human or a bot. Several DNN based systems have been achieved to fail this validation process. As a result, the reliability of CAPTCHA has become controversial. Adversarial image is a sample that is intentionally modified by adding non-random noise to deceive DNNs, and recently has been used to deceive DNNs to make CAPTCHA system more reliable. But the adversarial process degrades the image quality which makes it also difficult for humans to classify images and recognize objects. The proposed method preserves perceptual quality of adversarial images and still generates adversarial images which have 100% attack accuracy.

Date: 02.09.2019 / 10:00 Place: A-212

Investigation of the Theory of Mind was conducted in a verbal setting. This setting is constructed with help of some cognitive propositional attitude verbs (bil, düşün, and zannet) because they are used for referring to mental states as ToM is used for inferring mental states. In addition to the evaluation of verbs, participants’ ability to process pragmatic inferences were measured and these two sets of data were investigated for any relationship between them.

Date: 09.09.2019 / 10:00 Place: A-108

In this study, we presented a regression model, which was applied on cytotoxic bioactivity data obtained from HCC cells (from CanSyL dataset). Our objective was to predict off-target effects as potential new targets by regularizing the regression space based on the kinome tree topology. Our model was tested on the CanSyL dataset by applying LOOCV and achieved promising predictions. Then we scaled up our approach to the public datasets (CCLE and GDSC). Some kinase inhibitors were identified as outliers based on their individual RMSE. This difference suggests that outlier inhibitors are more specific inhibitors while non-outlier inhibitors are mostly multi-kinase inhibitors.

Date: 06.09.2019 / 10:00 Place: A-108

In the first part of this thesis, a BIM based architectural visualization and design tool is developed for VR, MR and PC environments. BIM provides detailed information and tools to professionals so that they can develop and manage buildings and infrastructures efficiently. In the second part of this thesis, two tutorial-like serious games, based on VR and MR environments, are developed to improve the users’ overall experience and to ease the user interactions in virtual environments. The developed tool and the serious games were tested by architects and game developers, and were evaluated from presence, usability, and technology acceptance perspectives.

Date: 29.08.2019 / 16:00 Place: A-212

Features extracted from speech are widely used for problems such as biometric speaker identification, but the use of speech data raises concerns about privacy. We propose a method for speech and posture classification using only breath data. The acoustical information was extracted from breath instances using the Hilbert-Huang transform and fed into our CNN-LSTM network for classification. We also created our publicly available dataset, BreathBase, which contains more than 4600 breath instances of 20 participants in 5 different postures with 4 different microphones. Using this data, 56% speaker classification and 96% posture classification accuracy is obtained.

Date: 02.09.2019 / 11:00 Place: A-212