Josip Job
Josip Job has a total of 12 co-authors. Here are the top co-authors by number of shared publications:
- Filip Bajić
University Computing Centre, University of Zagreb, 10000 Zagreb, Croatia
3 shared publications - Tomislav Horvat
Department of Electrical Engineering, University North, 104. Brigade 3, 42000 Varazdin, Croatia, EU
2 shared publications - Andrej Bošnjak
Faculty of Electrical Engineering, Computer Science and Information Technology Osijek, 31000 Osijek, Croatia
1 shared publication - Boris Lukić
Faculty of Agrobiotechnical Sciences Osijek, University of Josip Juraj Strossmayer of Osijek, 31000 Osijek, Croatia
1 shared publication - Denis Vranjes
Faculty of Electrical Engineering, Computer Science and Information Technology, J.J. Strossmayer University of Osijek, 31000 Osijek, Croatia
1 shared publication - Emmanuel Karlo Nyarko
Faculty of Electrical Engineering, Computer Science and Information Technology Osijek, Josip Juraj Strossmayer University of Osijek, Kneza Trpimira 2b, 31000 Osijek, Croatia
1 shared publication
Latest Publications
- Open Access
Validating Foundation Models for Automated Cattle Detection
Automated monitoring of cattle behavior through computer vision requires robust detection as a foundational step for tracking, re-identification, and behavior analysis. However, training accurate detection models typically demands extensive manually annotated datasets, creating a significant bottleneck for scaling cattle monitoring systems. In this work, we investigate whether automated annotation using foundation models can reliably replace manual labeling for cattle detection tasks. We introduce EMA (Extensive Mitrovac Annotations), a dataset of barn images with manually annotated cows with oriented bounding boxes including head orientation, posture labels (standing/lying), and visibility status (whole/partially visible). We systematically compare manualy annotated oriented bounding boxes with those generated by the Segment Anything Model 3 (SAM 3), demonstrating high agreement between automated and ground truth annotations. Furthermore, we train YOLO11-OBB detectors on both manual and SAM-generated annotations, showing that models trained on automated annotations achieve comparable performance to those trained on manual labels when evaluated on our ground truth test set. Our analysis reveals that only a small fraction of SAM-annotated data is sufficient to achieve high detection accuracy, proving the feasibility of automated annotation at scale. These findings suggest that foundation models show promise for generating training data in cattle detection pipelines under controlled conditions, potentially reducing annotation costs and supporting scalable deployment of monitoring systems. The EMA dataset and trained models are publicly available to support further research in precision livestock farming.
Open AccessValidating Foundation Models for Automated Cattle Detection
Automated monitoring of cattle behavior through computer vision requires robust detection as a foundational step for tracking, re-identification, and behavior analysis. However, training accurate detection models typically demands extensive manually annotated datasets, creating a significant bottleneck for scaling cattle monitoring systems. In this work, we investigate whether automated annotation using foundation models can reliably replace manual labeling for cattle detection tasks. We introduce EMA (Extensive Mitrovac Annotations), a dataset of barn images with manually annotated cows with oriented bounding boxes including head orientation, posture labels (standing/lying), and visibility status (whole/partially visible). We systematically compare manualy annotated oriented bounding boxes with those generated by the Segment Anything Model 3 (SAM 3), demonstrating high agreement between automated and ground truth annotations. Furthermore, we train YOLO11-OBB detectors on both manual and SAM-generated annotations, showing that models trained on automated annotations achieve comparable performance to those trained on manual labels when evaluated on our ground truth test set. Our analysis reveals that only a small fraction of SAM-annotated data is sufficient to achieve high detection accuracy, proving the feasibility of automated annotation at scale. These findings suggest that foundation models show promise for generating training data in cattle detection pipelines under controlled conditions, potentially reducing annotation costs and supporting scalable deployment of monitoring systems. The EMA dataset and trained models are publicly available to support further research in precision livestock farming.
- Open Access
A Data-Driven Machine Learning Algorithm for Predicting the Outcomes of NBA Games
We propose a new, data-driven model for the prediction of the outcomes of NBA and possibly other basketball league games by using machine learning methods. The paper starts with a strict mathematical formulation of the basketball statistical quantities and the performance indicators derived from them. The backbone of our model is the extended team efficiency index, which consists of two asymmetric parts: (i) the team efficiency index, generally based on some individual efficiency index—in our case, the NBA player efficiency index, and (ii) the comparing part, in which the observed team is rewarded for every selected feature in which it outperforms its rival. Based on the average of the past extended indices, the predicted extended indices are calculated symmetrically for both teams competing in the observed future game. The relative value of those indices defines the win function, which predicts the game outcome. The prediction model includes the concept of the optimal time window (OTW) for the training data. The training datasets were extracted from maximally four and the testing datasets from maximally two of the five consecutive observed NBA seasons (2013/2014–2017/2018). The model uses basic, derived, advanced, and league-wise basketball game elements as its features, whose preparation and extraction were briefly discussed. The proposed model was tested for several choices of the training and testing sets’ seasons, without and with OTWs. The average obtained prediction accuracy is around 66%, and the maximal obtained accuracy is around 78%. This is satisfactory and in the range of better results in the works of other authors.
Open AccessA Data-Driven Machine Learning Algorithm for Predicting the Outcomes of NBA Games
We propose a new, data-driven model for the prediction of the outcomes of NBA and possibly other basketball league games by using machine learning methods. The paper starts with a strict mathematical formulation of the basketball statistical quantities and the performance indicators derived from them. The backbone of our model is the extended team efficiency index, which consists of two asymmetric parts: (i) the team efficiency index, generally based on some individual efficiency index—in our case, the NBA player efficiency index, and (ii) the comparing part, in which the observed team is rewarded for every selected feature in which it outperforms its rival. Based on the average of the past extended indices, the predicted extended indices are calculated symmetrically for both teams competing in the observed future game. The relative value of those indices defines the win function, which predicts the game outcome. The prediction model includes the concept of the optimal time window (OTW) for the training data. The training datasets were extracted from maximally four and the testing datasets from maximally two of the five consecutive observed NBA seasons (2013/2014–2017/2018). The model uses basic, derived, advanced, and league-wise basketball game elements as its features, whose preparation and extraction were briefly discussed. The proposed model was tested for several choices of the training and testing sets’ seasons, without and with OTWs. The average obtained prediction accuracy is around 66%, and the maximal obtained accuracy is around 78%. This is satisfactory and in the range of better results in the works of other authors.
- 12 January 2023
- Published by Springer Nature in International Journal on Document Analysis and Recognition (IJDAR)
- Vol. 26 (4) , 453-474
Review of chart image detection and classification
- 12 January 2023
- Published by Springer Nature in International Journal on Document Analysis and Recognition (IJDAR)
- Vol. 26 (4) , 453-474
Review of chart image detection and classification
- Open Access
- 12 May 2022
- Published by MDPI AG in Journal of Imaging
- Vol. 8 (5) , 136
Data Extraction of Circular-Shaped and Grid-like Chart Images
Chart data extraction is a crucial research field in recovering information from chart images. With the recent rise in image processing and computer vision algorithms, researchers presented various approaches to tackle this problem. Nevertheless, most of them use different datasets, often not publicly available to the research community. Therefore, the main focus of this research was to create a chart data extraction algorithm for circular-shaped and grid-like chart types, which will accelerate research in this field and allow uniform result comparison. A large-scale dataset is provided containing 120,000 chart images organized into 20 categories, with corresponding ground truth for each image. Through the undertaken extensive research and to the best of our knowledge, no other author reports the chart data extraction of the sunburst diagrams, heatmaps, and waffle charts. In this research, a new, fully automatic low-level algorithm is also presented that uses a raster image as input and generates an object-oriented structure of the chart of that image. The main novelty of the proposed approach is in chart processing on binary images instead of commonly used pixel counting techniques. The experiments were performed with a synthetic dataset and with real-world chart images. The obtained results demonstrate two things: First, a low-level bottom-up approach can be shared among different chart types. Second, the proposed algorithm achieves superior results on a synthetic dataset. The achieved average data extraction accuracy on the synthetic dataset can be considered state-of-the-art within multiple error rate groups.
- 12 May 2022
- Published by MDPI AG in Journal of Imaging
- Vol. 8 (5) , 136
Open AccessData Extraction of Circular-Shaped and Grid-like Chart Images
Chart data extraction is a crucial research field in recovering information from chart images. With the recent rise in image processing and computer vision algorithms, researchers presented various approaches to tackle this problem. Nevertheless, most of them use different datasets, often not publicly available to the research community. Therefore, the main focus of this research was to create a chart data extraction algorithm for circular-shaped and grid-like chart types, which will accelerate research in this field and allow uniform result comparison. A large-scale dataset is provided containing 120,000 chart images organized into 20 categories, with corresponding ground truth for each image. Through the undertaken extensive research and to the best of our knowledge, no other author reports the chart data extraction of the sunburst diagrams, heatmaps, and waffle charts. In this research, a new, fully automatic low-level algorithm is also presented that uses a raster image as input and generates an object-oriented structure of the chart of that image. The main novelty of the proposed approach is in chart processing on binary images instead of commonly used pixel counting techniques. The experiments were performed with a synthetic dataset and with real-world chart images. The obtained results demonstrate two things: First, a low-level bottom-up approach can be shared among different chart types. Second, the proposed algorithm achieves superior results on a synthetic dataset. The achieved average data extraction accuracy on the synthetic dataset can be considered state-of-the-art within multiple error rate groups.
- Open Access
- 21 October 2021
- Published by MDPI AG in Journal of Imaging
- Vol. 7 (11) , 220
Chart Classification Using Siamese CNN
In recovering information from the chart image, the first step should be chart type classification. Throughout history, many approaches have been used, and some of them achieve results better than others. The latest articles are using a Support Vector Machine (SVM) in combination with a Convolutional Neural Network (CNN), which achieve almost perfect results with the datasets of few thousand images per class. The datasets containing chart images are primarily synthetic and lack real-world examples. To overcome the problem of small datasets, to our knowledge, this is the first report of using Siamese CNN architecture for chart type classification. Multiple network architectures are tested, and the results of different dataset sizes are compared. The network verification is conducted using Few-shot learning (FSL). Many of described advantages of Siamese CNNs are shown in examples. In the end, we show that the Siamese CNN can work with one image per class, and a 100% average classification accuracy is achieved with 50 images per class, where the CNN achieves only average classification accuracy of 43% for the same dataset.
- 21 October 2021
- Published by MDPI AG in Journal of Imaging
- Vol. 7 (11) , 220
Open AccessChart Classification Using Siamese CNN
In recovering information from the chart image, the first step should be chart type classification. Throughout history, many approaches have been used, and some of them achieve results better than others. The latest articles are using a Support Vector Machine (SVM) in combination with a Convolutional Neural Network (CNN), which achieve almost perfect results with the datasets of few thousand images per class. The datasets containing chart images are primarily synthetic and lack real-world examples. To overcome the problem of small datasets, to our knowledge, this is the first report of using Siamese CNN architecture for chart type classification. Multiple network architectures are tested, and the results of different dataset sizes are compared. The network verification is conducted using Few-shot learning (FSL). Many of described advantages of Siamese CNNs are shown in examples. In the end, we show that the Siamese CNN can work with one image per class, and a 100% average classification accuracy is achieved with 50 images per class, where the CNN achieves only average classification accuracy of 43% for the same dataset.
- Top Cited Papers
- 30 June 2020
- Published by Wiley in WIREs Data Mining and Knowledge Discovery
- Vol. 10 (5)
The use of machine learning in sport outcome prediction: A review
The increase in the volume of structured and unstructured data related to more than just sport events leads to the development and increased use of techniques that extract information and employ machine‐learning algorithms in predicting process outcomes based on input but not necessarily output data. Taking sports into consideration, predicting outcomes, and extracting valuable information has become appealing not only to sports workers but also to the wider audience, particularly in the areas of team management and sports betting. The aim of this article is to review the existing machine learning (ML) algorithms in predicting sport outcomes. Over 100 papers were analyzed and only some of these papers were taken into consideration. Almost all of the analyzed papers use some sort of feature selection and feature extraction, most often prior to using the machine‐learning algorithm. As an evaluation method of ML algorithms, researchers, in most cases, use data segmentation with data being chronologically distributed. In addition to data segmentation, researchers also use the k‐cross‐evaluation method. Sport predictions are usually treated as a classification problem with one class being predicted and rare cases being predicted as numerical values. Mostly used ML models are neural networks using data segmentation.This article is categorized under: Technologies > Machine Learning Technologies > Prediction
- 30 June 2020
- Published by Wiley in WIREs Data Mining and Knowledge Discovery
- Vol. 10 (5)
Top Cited PapersThe use of machine learning in sport outcome prediction: A review
The increase in the volume of structured and unstructured data related to more than just sport events leads to the development and increased use of techniques that extract information and employ machine‐learning algorithms in predicting process outcomes based on input but not necessarily output data. Taking sports into consideration, predicting outcomes, and extracting valuable information has become appealing not only to sports workers but also to the wider audience, particularly in the areas of team management and sports betting. The aim of this article is to review the existing machine learning (ML) algorithms in predicting sport outcomes. Over 100 papers were analyzed and only some of these papers were taken into consideration. Almost all of the analyzed papers use some sort of feature selection and feature extraction, most often prior to using the machine‐learning algorithm. As an evaluation method of ML algorithms, researchers, in most cases, use data segmentation with data being chronologically distributed. In addition to data segmentation, researchers also use the k‐cross‐evaluation method. Sport predictions are usually treated as a classification problem with one class being predicted and rare cases being predicted as numerical values. Mostly used ML models are neural networks using data segmentation.This article is categorized under: Technologies > Machine Learning Technologies > Prediction
- 14 July 2018
- Published by Taylor & Francis in International Journal of Electronics
- Vol. 106 (1) , 160-172
Chaos-Based transmitted-reference ultra-wideband communications
Transmitted-reference (TR) ultra-wideband (UWB) communication systems have gained increasing popularity for the usage in the low data rate application, due to its non-coherent receiver structure. In conventional TR system, non-coherency at the receiver is achieved by sending reference pulses prior to the data-bearing pulses. Then, at the receiver side, reference pulses are used as template signals for correlation with data-bearing pulses. Therefore, the orthogonality between reference and data pulses is obtained in time division multiple access (TDMA) fashion. However, the implementation of a wideband delay line is very difficult in current low power integrated circuits. In this paper, a TR method called Chaos-Based TR (CB-TR) is proposed. In the proposed method, chaotic sequences are used to separate the reference and data pulses. Such approach exploits the benefits of chaotic signals, such as non-periodicity, easy-to-generate, impulse-like autocorrelation value and low cross-correlation value. Furthermore, in order to decrease the influence of some negative properties of conventional chaotic maps, a modified chaotic generator (MCS) is proposed. Simulation results over the IEEE 802.15.4a channel model show comparable bit error rate performance to other TR methods.
- 14 July 2018
- Published by Taylor & Francis in International Journal of Electronics
- Vol. 106 (1) , 160-172
Chaos-Based transmitted-reference ultra-wideband communications
Transmitted-reference (TR) ultra-wideband (UWB) communication systems have gained increasing popularity for the usage in the low data rate application, due to its non-coherent receiver structure. In conventional TR system, non-coherency at the receiver is achieved by sending reference pulses prior to the data-bearing pulses. Then, at the receiver side, reference pulses are used as template signals for correlation with data-bearing pulses. Therefore, the orthogonality between reference and data pulses is obtained in time division multiple access (TDMA) fashion. However, the implementation of a wideband delay line is very difficult in current low power integrated circuits. In this paper, a TR method called Chaos-Based TR (CB-TR) is proposed. In the proposed method, chaotic sequences are used to separate the reference and data pulses. Such approach exploits the benefits of chaotic signals, such as non-periodicity, easy-to-generate, impulse-like autocorrelation value and low cross-correlation value. Furthermore, in order to decrease the influence of some negative properties of conventional chaotic maps, a modified chaotic generator (MCS) is proposed. Simulation results over the IEEE 802.15.4a channel model show comparable bit error rate performance to other TR methods.