HUCVL at SIU 2016

From May 16–19, Hacettepe University Computer Vision Lab (HUCVL) will be attending 24th IEEE Conference on Signal Processing and Communications Applications (SIU 2016), a major national annual conference on signal processing, pattern recognition and computer vision.

At the conference, we will be giving a short tutorial on deep learning, which will be on May 16th between 11:00 – 12:30. Moreover, we will be presenting 6 oral papers, some of which are chosen as runner-ups for Alper Atalay Best Student Paper Award. The descriptions of these papers can be found below.

See you all at Zonguldak in two weeks time…

Clustering Motion Trajectories via Dominant Sets (Runner-Up for Alper Atalay Best Student Paper Award)

Authors: Cagdas Bak, Aykut Erdem, Erkut Erdem
Abstract — In recent years, video surveillance systems stand out as an important research topic in the field of computer vision. Studies in this context usually focus on detecting common motion patterns in video sequences, determining unexpected motions or predicting possible future events. Performance of these studies directly depends on the performances of the pre-processing steps. In this paper, we present an approach to extract motion trajectories and then to cluster those trajectories that have similar characteristics. The proposed approach can be considered as a preliminary step for complex surveillance systems. Unlike the existing studies in the literature, the proposed method both produces more accurate results for determining common motion patterns and yields successful results on long video sequences in consequence of used clustering algorithm. The effectiveness and the performance of the proposed approach is validated on VIRAT dataset.

Dominant Sets Based Analysis of Human Crowds

Authors: Burcak Asal, Aykut Erdem, Erkut Erdem
Abstract — Due to recent advances in new camera technologies and the Internet, millions of videos can be easily accessed from any place at any time. A significant amount of these videos are for surveillance, and include actors such as humans and vehicles performing different actions in dynamic scenes. The goal of this study is to analyze human crowd motions in videos. More specifically, moving humans are tracked throughout a video sequence, and the collective crowd motions are then clustered using path similarities via the Dominant Sets method. Moreover, this clustering result can be used to predict the coherency of the motion as a scalar value.

Summarizing Personal Image Collections with Intrinsic Properties (Runner-Up for Alper Atalay Best Student Paper Award)

Authors: Goksu Erdogan, Bora Celikkale, Aykut Erdem, Erkut Erdem
Abstract — Visual summarization problems are complex intrinsically. Because definition of a summary have some ambiguity and only one correct summary does not exist. A good summary consists of two main properties in general which are (i) coverage and (ii) diversity. A good summary should have high coverage. On the other words summary should consist of key events and concepts for given set. At the same time a good summary should also be diverse, i.e it should not consist of similar events and concepts. In addition to these two main properties, intrinsic properties such as quality, emotions, popularity of images increase in importance depending on prevalence of social media applications. We proposed an automatic summarization method which considers intrinsic properties of images in addition to coverage and diversity for personal image collection summarization. This developed method is compared with summaries of different personal image collections which are collected by crowdsourcing and it is observed that the taking intrinsic properties into account improves the summaries.

TasvirEt: A Benchmark Dataset for Automatic Turkish Description Generation from Images (Runner-Up for Alper Atalay Best Student Paper Award)

Authors: Mesut Erhan Unal, Begum Citamak, Semih Yagcioglu, Aykut Erdem, Erkut Erdem, Nazli Ikizler Cinbis, Ruket Cakici
Abstract — Automatically describing images with natural sentences is considered to be a challenging research problem that has recently been explored. Although the number of methods proposed to solve this problem increases over time, since the datasets used commonly in this field contain only English descriptions, the studies have mostly been limited to single language, namely English. In this study,for the first time in the literature, a new dataset is proposed which enables generating Turkish descriptions from images, which can be used as a benchmark for this purpose. Furthermore, two approaches are proposed, again for the first time in the literature, for image captioning in Turkish with the dataset we named as TasvirEt. Our findings indicate that the new Turkish dataset and the approaches used here can be successfully used for automatically describing images in Turkish.

Turkish Sign Language Recognition Application Using Motion History Image

Authors: Ozge Yalcinkaya, Anil Atvar, Pinar Duygulu
Abstract — Recognizing sign language is an important interest area since there are many speech and hearing impaired people in the world. They need to be understood by other people and understand them as well. Unfortunately, the number of people who have the knowledge of sign language is not many. In order to communicate with handicapped people, existence of some automatized systems may be helpful. Therefore, in this work, we aimed to implement a system that recognizes the sign language and converts it to text to help people while communicating with each other where the input scene is taken from camera. We produced a training data which includes eight different sign language videos. After that, we used “Motion History Images”(MHI) to extract the motion information from them. A classification is done by using nearest neighbor approach after extracting the features from MHI of videos. As a result, by using training data, our system predicts the text for given sign language. The overall classification accuracy is computed as 95%.

Visual Saliency Estimation via Attribute Based Classifiers and Conditional Random Field (Runner-Up for Alper Atalay Best Student Paper Award)

Authors: Berkan Demirel, R. Gokberk Cinbis, Nazli Ikizler-Cinbis
Abstract — Visual Saliency Estimation is a computer vision problem that aims to find the regions of interest that are frequently in eye focus in a scene or an image. Since most computer vision problems require discarding irrelevant regions in a scene, visual saliency estimation can be used as a preprocessing step in such problems. In this work, we propose a method to solve top-down saliency estimation problem using Attribute Based Classifiers and Conditional Random Fields (CRF). Experimental results show that attribute-based classifiers encode visual information better than low level features and the presented approach generates promising results compared to state-of-the-art approaches on Graz-02 dataset.