Araştırma Çıktıları | WoS | Scopus | TR-Dizin | PubMed
Permanent URI for this communityhttps://hdl.handle.net/20.500.14719/1741
Browse
7 results
Search Results
Publication Metadata only Creation of Annotated Synthetic UAV Video Dataset for Object Detection and Tracking(IEEE, 2023) Yilmaz, Can; Maras, Bahri; Arica, Nafiz; Ertuzun, Aysin Baytan; Bahcesehir University; Bogazici University; Piri Reis UniversityIn order for object detection and tracking in videos obtained from unmanned aerial vehicles (UAVs) by deep convolutional neural networks (DCNN), extensive ground truth optical flow, occlusion and segmentation datasets, of various objects or vehicles, are required during the training and testing processes. The mentioned ground truth informations are not widely available in the literature due to the difficulty of labeling or extracting them from real-life recorded UAV video images. In this study, ground truth optical flow, occlusion and segmentation datasets were produced synthetically for the first time with the UAV point of view in a novel way, so as to fill the gap in literature. The ground truth datasets were created for each vehicle by subjecting the triangles (mesh) automatically generated by the Unity engine to the homography method. With this method, 1920x1080 and 250x250 sized synthetic datasets consisting of 100 scenarios were obtained.Publication Metadata only Enhancing Object Detection Algorithms by Synthetic Aerial Images(IEEE, 2023) Yilmaz, Can; Maras, Bahri; Yilmaz, Gorkem; Ceylan, Goksu; Hamamcioglu, Onder; Arica, Nafiz; Ertuzun, Aysin Baytan; Bahcesehir University; Bogazici University; Bahcesehir University; Piri Reis UniversityIn order to accurately perform object detection by deep convolutional neural networks (DCNN) in videos, obtained from unmanned aerial vehicles (UAVs), many example images of objects containing annotations such as ground truth class information, bounding box, optical flow, occlusion and segmentation are required. Due to the difficulties faced during annotation of scenarios, and due to the inadequacy of the scenario diversity resulting from environmental conditions, a dataset containing above mentioned ground truths has not been found in the literature. In this study, synthetic aerial images with various annotation information were created in different scenarios while composing virtual worlds, and enhancing object detection algorithms is aimed. Enhancement of detection results of DCNN based object detection algorithms, trained with the support of synthetic aerial images, on real-world aerial images significantly, was observed during the experiments, conducted.Publication Metadata only A Scoring Method for Interpretability of Concepts in Convolutional Neural Networks(IEEE, 2022) Gurkan, Mustafa Kagan; Arica, Nafiz; Vural, Fato Yarman; Bahcesehir University; Middle East Technical UniversityIn this paper, we propose a scoring algorithm for measuring the interpretability of CNN models by focusing on the feature extraction operation at the convolutional layers. The proposed approach is based on the principal of concept analysis, for a predefined list of concepts. A map of the network is created based on its responsiveness against each concept. Once this map is ready, various images can be applied as inputs and they are matched with the concepts whose hidden nodes are highly activated. Finally, the evaluation algorithm kicks in to use these descriptions during the final prediction and provides human-understandable explanations.Publication Metadata only Graph Neural Networks Based Approach for Interpersonal Relationship Classification in Images(IEEE, 2023) Akay, Simge; Arica, Nafiz; Bahcesehir University; Piri Reis UniversityInterpersonal relationships, which are an integral part of our social life, demonstrate how people connect and interact within society. Describing the relationship between two individuals in images depends on many different factors and attributes. In this study, a Graph Neural Network (GNN) based approach is proposed that focuses on the important attributes for describing the relationship between two individuals in images. In the proposed method, each attribute that will be used to describe the relationship is defined as a GNN node. Then, the meaningless connections between nodes are pruned with a pruning operation to obtain the ideal GNN model. In this study, a more robust GNN for classifying interpersonal relationships is created by using rich attributes, unlike the literature, and obtaining significant connections between nodes through pruning. The experiments conducted in this study showed that both using a wider GNN and pruning operation improve the classification performance.Publication Metadata only Microservice Interaction Prediction in Communication Platform as a Service(IEEE, 2022) Aktas, Kemal; Kilinc, H. Hakan; Arica, Nafiz; Bahcesehir UniversityIn telecommunication platforms, it is necessary to monitor the system, take quick action against possible and observed errors, and ensure the continuous operability of the system. However, debugging and problem addressing processes can take a long time in microservice architecture-based platforms with high number of users and heavy traffic. This difficulty arises from the necessity of detecting microservice interactions. In the first phase of this study, it has been observed that providing microservice flows and interactions saves operational teams a significant time during the debugging process. In this scope, different machine learning-based models that predict microservice interactions have been developed and performances are compared. In these models, log patterns are extracted on microservice log data and the interaction map of the mentioned microservice is created by estimating the previous and next microservices that the current microservice interacts with at a certain moment. In the experiments on microservice logs working with the basic call scenario, successful estimation results were obtained that could contribute positively to the debugging process.Publication Metadata only Real-Time Image Stitching For Multiple Camera Panoramic Video Shoot: A Case Study in Football Matches(IEEE, 2020) Bayrak, Mehmet; Kilinc, Orkun; Arica, Nafiz; Bahcesehir UniversityIn this study, a real-time image stitching method is proposed for cost-effective high resolution wide angle video shooting. In the first stage, the images taken from more than one camera with fixed position are stitched with a classical algorithm. After the parameters calculated in the first stage are stored, the image pixels are mapped using the stored parameters and ArUco markers. The mapping process is used for real-time panoramic video shooting after it has been calculated for that particular camera setup once. The images taken from the multi camera are combined by remapping with a GPU based approach. Therefore, registered mapping can be used in different environments without changing the position and lenses of the cameras. As a case study, real-time panoramic video is shot with two cost-effective cameras in football matches. Deep-learning based autonomous pilot video shooting is then performed on the high resolution panoramic video obtained. In experiments, 36 FPS speed has been reached by using a standard desktop computer and it has been seen that image quality measurements are at reasonable levels.Publication Metadata only Effects of Network Depths on Semantic Image Segmentation By Weakly Supervised Learning(IEEE, 2020) Bircanoglu, Cenk; Arica, Nafiz; Bahcesehir UniversityWeakly Supervised Learning is one of the most interesting approaches that more complex labels are predicted by using related simple labels. In this study, we focus on segmentation problem by giving image class tags in learning stage. We examine how the number of layers and the usage of their output in Convolutional Neural Network affect the segmentation results. It is found that increasing the number of layers in the network has a positive effect on segmentation performance. After ResNet152 is determined as the most successful deep architecture in Pascal VOC2012 dataset, we construct a new architecture based on ResNet152. Experimental results show that proposed architecture outperforms the available studies tested on this particular dataset. In addition, we observe that early layers reach more general attributes for the object classes than the last layers and that these attributes can better identify the object boundaries.