Computer Engineering / Bilgisayar Mühendisliği
Permanent URI for this collectionhttps://hdl.handle.net/11147/10
Browse
5 results
Search Results
Conference Object Citation - WoS: 7Citation - Scopus: 6Semantic Pose Verification for Outdoor Visual Localization With Self-Supervised Contrastive Learning(IEEE, 2022) Guerrero, Jose J.; Orhan, Semih; Baştanlar, YalınAny city-scale visual localization system has to overcome long-term appearance changes, such as varying illumination conditions or seasonal changes between query and database images. Since semantic content is more robust to such changes, we exploit semantic information to improve visual localization. In our scenario, the database consists of gnomonic views generated from panoramic images (e.g. Google Street View) and query images are collected with a standard field-of-view camera at a different time. To improve localization, we check the semantic similarity between query and database images, which is not trivial since the position and viewpoint of the cameras do not exactly match. To learn similarity, we propose training a CNN in a self-supervised fashion with contrastive learning on a dataset of semantically segmented images. With experiments we showed that this semantic similarity estimation approach works better than measuring the similarity at pixel-level. Finally, we used the semantic similarity scores to verify the retrievals obtained by a state-of-the-art visual localization method and observed that contrastive learning-based pose verification increases top-1 recall value to 0.90 which corresponds to a 2% improvement.Conference Object Citation - WoS: 5Citation - Scopus: 10Efficient Search in a Panoramic Image Database for Long-Term Visual Localization(IEEE, 2021) Orhan, Semih; Baştanlar, YalınIn this work, we focus on a localization technique that is based on image retrieval. In this technique, database images are kept with GPS coordinates and the geographic location of the retrieved database image serves as an approximate position of the query image. In our scenario, database consists of panoramic images (e.g. Google Street View) and query images are collected with a standard field-of-view camera in a different time. While searching the match of a perspective query image in a panoramic image database, unlike previous studies, we do not generate a number of perspective images from the panoramic image. Instead, taking advantage of CNNs, we slide a search window in the last convolutional layer belonging to the panoramic image and compute the similarity with the descriptor extracted from the query image. In this way, more locations are visited in less amount of time. We conducted experiments with state-of-the-art descriptors and results reveal that the proposed sliding window approach reaches higher accuracy than generating 4 or 8 perspective images.Article Citation - WoS: 43Citation - Scopus: 47Semantic Segmentation of Outdoor Panoramic Images(Springer, 2021) Orhan, Semih; Baştanlar, YalınOmnidirectional cameras are capable of providing 360. field-of-view in a single shot. This comprehensive view makes them preferable for many computer vision applications. An omnidirectional view is generally represented as a panoramic image with equirectangular projection, which suffers from distortions. Thus, standard camera approaches should be mathematically modified to be used effectively with panoramic images. In this work, we built a semantic segmentation CNN model that handles distortions in panoramic images using equirectangular convolutions. The proposed model, we call it UNet-equiconv, outperforms an equivalent CNN model with standard convolutions. To the best of our knowledge, ours is the first work on the semantic segmentation of real outdoor panoramic images. Experiment results reveal that using a distortion-aware CNN with equirectangular convolution increases the semantic segmentation performance (4% increase in mIoU). We also released a pixel-level annotated outdoor panoramic image dataset which can be used for various computer vision applications such as autonomous driving and visual localization. Source code of the project and the dataset were made available at the project page (https://github.com/semihorhan/semseg-outdoor-pano). © 2021, The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature.Article Citation - WoS: 9Citation - Scopus: 13Training Cnns With Image Patches for Object Localisation(Institution of Engineering and Technology, 2018) Orhan, Semih; Baştanlar, YalınRecently, convolutional neural networks (CNNs) have shown great performance in different problems of computer vision including object detection and localisation. A novel training approach is proposed for CNNs to localise some animal species whose bodies have distinctive patterns such as leopards and zebras. To learn characteristic patterns, small patches which are taken from different body parts of animals are used to train models. To find object location, in a test image, all locations are visited in a sliding window fashion. Crops are fed into trained CNN and their classification scores are combined into a heat map. Later on, heat maps are converted to bounding box estimates for varying confidence scores. The localisation performance of the patch-based training approach is compared with Faster R-CNN – a state-of-the-art CNN-based object detection and localisation method. Experimental results reveal that the patch-based training outperforms Faster R-CNN, especially for classes with distinctive patterns.Conference Object Parça Tabanlı Eǧitimin Evrişimli Yapay Sinir Aǧları ile Nesne Konumlandırma Üzerindeki Etkisi(IEEE, 2017) Orhan, Semih; Bastanlar, YalinIn recent years, Convolutional Neural Networks (CNNs) have shown great performance not only in image classification and image recognition tasks but also several tasks of computer vision. A lot of models which have different number of layers and depths, have been proposed. In this work, locations of leopards are tried to be identified by deep neural networks. To accomplish this task, two different methods are applied. First of them is training neural network using with entire images, second of them is training neural networks using with image patches which are cropped from full size of images. Patch training model has shown better performance than full size of image trained model.