No abstract

Confusion matrix is a useful and comprehensive presentation of the classifier performance. It is commonly used in the evaluation of multi-class, single-label classification models, where each data instance can belong to just one class at any given point in time. However, the real world is rarely unambiguous and hard classification of data instance to a single class, i.e. defining its properties with single distinctive feature, is not always possible. For example, an image can contain multiple objects and regions which makes multi-class classification inappropriate to describe its content. Proposed solutions to this set of problems are based on multi-label classification model where each data instance is assigned one or more labels describing its features. While most of the evaluation measures used to evaluate single-label classifier can be adapted to a multi-label classification model, presentation and evaluation of the obtained results using standard confusion matrices cannot be expanded to this case.In this paper we propose a novel method for the computation of a confusion matrix for multi-label classification. The proposed algorithm overcomes the limitations of the existing approaches in modeling relations between the classifier output and the Ground Truth (i.e. hand-labeled) classification, and due to its versatility can be used in many different research fields.

In this paper, a novel approach for an automatic object detection and localisation on aerial images is proposed. Proposed model does not use ground control points (GCPs) and consists of three major phases. In the first phase, optimal flight route is planned in order to capture the area of interest and aerial images are acquired using unmanned aerial vehicle (UAV), followed by creating a mosaic of collected images to obtained larger field-of-view panoramic image of the area of interest and using the obtained image mosaic to create georeferenced map. The image mosaic is then also used to detect objects of interest using the approach based on convolutional neural networks.

Unmanned Aircraft Systems (UASs) have been recognized as an important resource in search-and-rescue (SAR) missions and, as such, have been used by the Croatian Mountain Search and Rescue (CMRS) service for over seven years. The UAS scans and photographs the terrain. The high-resolution images are afterwards analyzed by SAR members to detect missing persons or to find some usable trace. It is a drawn out, tiresome process prone to human error. To facilitate and speed up mission image processing and increase detection accuracy, we have developed several image-processing algorithms. The latest are convolutional neural network (CNN)-based. CNNs were trained on a specially developed image database, named HERIDAL. Although these algorithms achieve excellent recall, the efficiency of the algorithm in actual SAR missions and its comparison with expert detection must be investigated. A series of mission simulations are planned and recorded for this purpose. They are processed and labelled by a developed algorithm. A web application was developed by which experts analyzed raw and processed mission images. The algorithm achieved better recall compared to an expert, but the experts achieved better accuracy when they analyzed images that were already processed and labelled.

In recent years Generative Adversarial Networks (GANs) have achieved remarkable results in the task of realistic image synthesis. Despite their continued success and advances, there still lacks a thorough understanding of how precisely GANs map random latent vectors to realistic-looking images and how the priors set on the latent space affect the learned mapping. In this work, we analyze the effect of the chosen latent dimension on the final quality of synthesized images of human faces and learned data representations. We show that GANs can generate images plausibly even with latent dimensions significantly smaller than the standard dimensions like 100 or 512. Although one might expect that larger latent dimensions encourage the generation of more diverse and enhanced quality images, we show that an increase of latent dimension after some point does not lead to visible improvements in perceptual image quality nor in quantitative estimates of its generalization abilities.