Within the last decade, neural network based predictors have demonstrated impressive - and at times superhuman capabilities. This performance is often paid for with an intransparent prediction process thus has sparked numerous contributions in novel field of explainable artificial intelligence (XAI). In this paper, we focus on a popular widely used method XAI, Layer-wise Relevance Propagation (LRP). Since its initial proposition LRP evolved as method, best practice applying tacitly emerged,...

The task of structured output prediction deals with learning general functional dependencies between arbitrary input and spaces. In this context, two loss-sensitive formulations for maximum-margin training have been proposed in the literature, which are referred to as margin slack rescaling, respectively. latter is believed be more accurate easier handle. Nevertheless, it not popular due lack known efficient inference algorithms; therefore, rescaling--which requires a similar type normal...

We focus on detecting anomalies in images where the data distribution is supported by a lower-dimensional embedded manifold. Approaches based autoencoders have aimed to control their capacity either reducing size of bottleneck layer or imposing sparsity constraints activations. However, none these techniques explicitly penalize reconstruction anomalous regions, often resulting poor detection. tackle this problem adapting self-supervised learning regime that essentially implements denoising...

The task of natural language parsing can naturally be embedded in the maximum-margin framework for structured output prediction using an appropriate joint feature map and a suitable loss function. While there are efficient learning algorithms based on cutting-plane method optimizing resulting quadratic objective with potentially exponential number linear constraints, their efficiency crucially depends inference used to infer most violated constraint current iteration. In this paper, we...

Real-time object tracking, feature assessment and classification based on video are an enabling technology for improving situation awareness of human operators as well automated recognition critical situations. To bridge the gap between signal-processing output spatio-temporal analysis behavior at semantic level, a generic sensor-independent representation is necessary. However, in case public corporate surveillance, centralized storage aggregated data leads to privacy violations. This...

The interpretation of complex scenes requires a large amount prior knowledge and experience. To utilize in computer vision or decision support system for image interpretation, probabilistic scene model is developed. In conjunction with the observer's characteristics (a human interpreter system), it possible to bottom-up inference from observations as well focus attention observer on most promising classes objects. presented Bayesian approach allows rigorous formulation uncertainty models...

Structural support vector machine (SVM) is an elegant approach for building complex and accurate models with structured outputs. However, its applicability relies on the availability of efficient inference algorithms-the state-of-the-art training algorithms repeatedly perform to compute a subgradient or find most violating configuration. In this paper, we propose exact algorithm maximizing nondecomposable objectives due special type high-order potential having decomposable internal...

From the advances in computer vision methods for detection, tracking and recognition of objects video streams, new opportunities surveillance arise: In future, automated systems will be able to detect critical situations early enough enable an operator take preventive actions, instead using material merely forensic investigations. However, problems such as limited computational resources, privacy regulations a constant change potential threads have addressed by practical system. this paper,...

The evaluation of a country's critical infrastructure requires detailed analysis facilities such as airfields, harbors and heavy industry. To improve the assessment facilities, an assistance system for interpretation from aerial imagery is developed. In this paper we point out recent advances system's recommendation function. Besides suggesting occurrence undetected objects based on probabilistic scene model previously detected objects, now able to suggest classification intrinsic object...

The evaluation of a country's critical infrastructure requires detailed analysis facilities such as airfields, harbors, communication lines and heavy industry. To improve the interpretation process, an interactive support system for from aerial imagery is developed. aim to facilitate training phase beginners, increase flexibility in assignment interpreters overall quality interpretation. An approach chosen by professional has been basis identify steps which can be effectively supported...

Many learning tasks in the field of natural language processing including sequence tagging, segmentation, and syntactic parsing have been successfully approached by means structured prediction methods. An appealing property corresponding training algorithms is their ability to integrate loss function interest into optimization process improving final results according chosen measure performance. Here, we focus on task constituency show how optimize model for F1 -score max-margin framework a...

Considering the worst-case scenario, junction-tree algorithm remains most general solution for exact MAP inference with polynomial run-time guarantees. Unfortunately, its main tractability assumption requires treewidth of a corresponding MRF to be bounded, strongly limiting range admissible applications. In fact, many practical problems in area structured prediction require modeling global dependencies by either directly introducing factors or enforcing constraints on variables. However,...

Object recognition is a typical task of aerial reconnaissance and especially in military applications, to determine the class an unknown object on battlefield can give valuable information its capabilities threat. RecceMan® (Reconnaissance Manual) decision support system for developed by Fraunhofer IOSB. It supports automating tedious matching features with set possible classes, while leaving assessment trained human interpreter. The quality assessed user influenced several factors such as...

Markov random fields (MRFs) are a powerful tool for modelling statistical dependencies set of variables using graphical representation. An important computational problem related to MRFs, called maximum posteriori (MAP) inference, is finding joint variable assignment with the maximal probability. It well known that two popular optimisation techniques this task, linear programming (LP) relaxation and dual decomposition (DD), have strong connection both providing an optimal solution MAP when...

The analysis of complex infrastructure from aerial imagery, for instance a detailed an airfield, requires the interpreter, besides to be familiar with sensor's imaging characteristics, have understanding domain. required domain knowledge includes about processes and functions involved in operation infrastructure, potential objects used provide those their spatial functional interrelations. Since it is not possible yet reliable automatic object recognition (AOR) such scenes, we developed...