Search

The advancement of deep learning methods across various applications has forced the creation of enormous training datasets. However, obtaining suitable real-world datasets is often challenging for various reasons. Consequently, numerous studies have emerged focusing on the generation and utilization of synthetic data in the training process. Hence, there is no universal formula for preparing synthetic data and leveraging it in network training to maximize the effectiveness of various detection methods. This work provides a comprehensive overview of several synthetic data generation techniques, followed by a thorough investigation into the impact of training methods and the selection of synthetic data quantities. The outcomes of this research enable the formulation of conclusions regarding the recipe for developing synthetic data with high efficacy in enhancing detection methods. The main conclusion for the synthetic data generation methods is to ensure maximum diversity at a high level of photorealism, which allows improving the classification quality by more than 5% to even 19% for different detection metrics. [ABSTRACT FROM AUTHOR]

The automatic detection of violent actions in public places through video analysis is difficult because the employed Artificial Intelligence-based techniques often suffer from generalization problems. Indeed, these algorithms hinge on large quantities of annotated data and usually experience a drastic drop in performance when used in scenarios never seen during the supervised learning phase. In this paper, we introduce and publicly release the Bus Violence benchmark, the first large-scale collection of video clips for violence detection on public transport, where some actors simulated violent actions inside a moving bus in changing conditions, such as the background or light. Moreover, we conduct a performance analysis of several state-of-the-art video violence detectors pre-trained with general violence detection databases on this newly established use case. The achieved moderate performances reveal the difficulties in generalizing from these popular methods, indicating the need to have this new collection of labeled data, beneficial for specializing them in this new scenario. [ABSTRACT FROM AUTHOR]

New diseases constantly endanger the lives of populations, and, nowadays, they can spread easily and constitute a global threat. The COVID-19 pandemic has shown that the fight against a new disease may be difficult, especially at the initial stage of the epidemic, when medical knowledge is not complete and the symptoms are ambiguous. The use of machine learning tools can help to filter out those sick patients who do not need to be tested for spreading the pathogen, especially in the event of an overwhelming increase in disease transmission. This work presents a screening support system that can precisely identify patients who do not carry the disease. The decision of the system is made on the basis of patient survey data that are easy to collect. A case study on a data set of symptomatic COVID-19 patients shows that the system can be effective in the initial phase of the epidemic. The case study presents an analysis of two classifiers that were tuned to achieve an assumed acceptable threshold of negative predictive values during classification. Additionally, an explanation of the obtained classification models is presented. The explanation enables the users to understand the basis of the decision made by the model. The obtained classification models provide the basis for the DECODE service, which could serve as support in screening patients with COVID-19 disease at the initial stage of the pandemic. Moreover, the data set constituting the basis for the analyses performed is made available to the research community. This data set, consisting of more than 3000 examples, is based on questionnaires collected at a hospital in Poland. [ABSTRACT FROM AUTHOR]

Efforts of the scientific community led to the development of multiple screening approaches for COVID-19 that rely on machine learning methods. However, there is a lack of works showing how to tune the classification models used for such a task and what the tuning effect is in terms of various classification quality measures. Understanding the impact of classifier tuning on the results obtained will allow the users to apply the provided tools consciously. Therefore, using a given screening test they will be able to choose the threshold value characterising the classifier that gives, for example, an acceptable balance between sensitivity and specificity. The presented work introduces the optimisation approach and the resulting classifiers obtained for various quality threshold assumptions. As a result of the research, an online service was created that makes the obtained models available and enables the verification of various solutions for different threshold values on new data. [ABSTRACT FROM AUTHOR]

Tracking and action-recognition algorithms are currently widely used in video surveillance, monitoring urban activities and in many other areas. Their development highly relies on benchmarking scenarios, which enable reliable evaluations/improvements of their efficiencies. Presently, benchmarking methods for tracking and action-recognition algorithms rely on manual annotation of video databases, prone to human errors, limited in size and time-consuming. Here, using gained experiences, an alternative benchmarking solution is presented, which employs methods and tools obtained from the computer-game domain to create simulated video data with automatic annotations. Presented approach highly outperforms existing solutions in the size of the data and variety of annotations possible to create. With proposed system, a potential user can generate a sequence of random images involving different times of day, weather conditions, and scenes for use in tracking evaluation. In the design of the proposed tool, the concept of crowd simulation is used and developed. The system is validated by comparisons to existing methods. [ABSTRACT FROM AUTHOR]