High Classification Scores Research Articles

Astrometric Binary Classification via Artificial Neural Networks

With nearly two billion stars observed and their corresponding astrometric parameters evaluated in the recent Gaia mission, the number of astrometric binary candidates has risen significantly. Due to the surplus of astrometric data, the current computational methods employed to inspect these astrometric binary candidates are both computationally expensive and cannot be executed in a reasonable time frame. In light of this, a machine learning (ML) technique to automatically classify whether a set of stars belongs to an astrometric binary pair via an artificial neural network (ANN) is proposed. Using data from Gaia Data Release 3, the ANN was trained and tested on 1.5 million highly probable true and visual binaries, considering the proper motions, parallaxes, and angular and physical separations as features. The ANN achieves high classification scores, with an accuracy of 99.3%, a precision rate of 0.988, a recall rate of 0.991, and an area under the curve of 0.999, indicating that the utilized ML technique is a highly effective method for classifying astrometric binaries. Thus, the proposed ANN is a promising alternative to the existing methods for the classification of astrometric binaries.

Enhancing trustworthiness in ML-based network intrusion detection with uncertainty quantification

A crucial role in the security of modern networks is played by Intrusion Detection Systems (IDSs), security devices designed to identify and mitigate attacks to networks structure. Data-driven approaches based on Machine Learning (ML) have gained more and more popularity for executing the classification tasks required by signature-based IDSs. However, typical ML models adopted for this purpose do not properly take into account the uncertainty associated with their prediction. This poses significant challenges, as they tend to produce misleadingly high classification scores for both misclassified inputs and inputs belonging to unknown classes (e.g. novel attacks), limiting the trustworthiness of existing ML-based solutions. In this paper, we argue that ML-based IDSs should always provide accurate uncertainty quantification to avoid overconfident predictions. In fact, an uncertainty-aware classification would be beneficial to enhance closed-set classification performance, would make it possible to carry out Active Learning, and would help recognize inputs of unknown classes as truly unknowns, unlocking open-set classification capabilities and Out-of-Distribution (OoD) detection. To verify it, we compare various ML-based methods for uncertainty quantification and open-set classification, either specifically designed for or tailored to the domain of network intrusion detection. Moreover, we develop a custom model based on Bayesian Neural Networks that stands out for its OoD detection capabilities and robustness, with a lower variance in the results over different scenarios, compared to other baselines, thus showing how proper uncertainty quantification can be exploited to significantly enhance the trustworthiness of ML-based IDSs.

In the Wild Video Violence Detection: An Unsupervised Domain Adaptation Approach

This work addresses the challenge of video violence detection in data-scarce scenarios, focusing on bridging the domain gap that often hinders the performance of deep learning models when applied to unseen domains. We present a novel unsupervised domain adaptation (UDA) scheme designed to effectively mitigate this gap by combining supervised learning in the train (source) domain with unlabeled test (target) data. We employ single-image classification and multiple instance learning (MIL) to select frames with the highest classification scores, and, upon this, we exploit UDA techniques to adapt the model to unlabeled target domains. We perform an extensive experimental evaluation, using general-context data as the source domain and target domain datasets collected in specific environments, such as violent/non-violent actions in hockey matches and public transport. The results demonstrate that our UDA pipeline substantially enhances model performances, improving their generalization capabilities in novel scenarios without requiring additional labeled data.

Classification of non-referenced continuous-wave terahertz reflection spectra for remote material identification

Commonly, terahertz spectra (both continuous-wave and pulsed) are deconvoluted by reference spectra to remove the water vapor absorption lines and other system related responses. However, in real-life applications obtaining reference spectra can be problematic and adds to the complexity of the system. Thus, a reference-free method for classification of terahertz spectra could be a welcomed advance for remote sensing applications. In this paper, we study how simple machine learning algorithms perform as a reference-free method for terahertz stand-off identification of materials. The algorithms are trained using spectra measured under controlled humidity conditions and tested by a completely independent data set measured under ambient conditions. We apply three different classification algorithms; namely a Gaussian Bayes model, the k nearest neighbors, and a support vector machine. We found that, if the terahertz spectra are processed using a supervised algorithm (Regularized Linear Discriminant Analysis), very high classification scores (¿98.6%) can be retained for the non-referenced spectra. Moreover, the high accuracy is obtained meanwhile the dimensionality is reduced by a factor larger than 160, which further reduces the computational requirements. Hence, we have demonstrated that simple supervised machine learning algorithms can serve as a highly accurate reference-free method for THz material identification. This could be of great importance for real-world remote sensing applications based on terahertz spectroscopy.

Association Between the Society for Vascular Surgery (SVS)-WIfI (Wound, Ischemia, Foot Infection) Classification, Wagner-Meggitt Classification, and Amputation Rate in Patients With Diabetic Foot Infection.

Diabetic foot infection (DFI) is a common problem in patients with diabetic foot disease. Amputations and other complications of DFI lead to significant morbidity and mortality. The Society for Vascular Surgery (SVS)-WIfI (wound, ischemia, and foot infection) classification system can evaluate the benefit from revascularization and the risk of amputation in 1 and 3 years. We aimed to evaluate SVS-WIfI and Wagner-Meggitt (WM) prediction of DFI outcome, and to determine factors associated with major amputation and mortality rate. The patients with diabetes who presented between June 2018 and May 2020 with characteristics suggesting a more serious DFI or potential indications for hospitalization were reviewed in this cohort study. Demographic data, clinical characteristics, and type of revascularization were evaluated. One-year and 3-year amputation and mortality rates were the main outcomes. The grading of WM classification and the SVS-WIfI score were compared between amputation and nonamputation groups. Association between mortality and comorbidity were analyzed. One hundred and thirty-one patients admitted with DFI were included in study. And 73.28% had peripheral arterial disease (PAD). The 1-year and 3-year major amputation rates were 16.03% and 26.23%, respectively. Seventy-eight (59.54%) patients required minor amputation to control infection before revascularization. PAD (risk ratio [RR] 1.47: 95% confidence interval [CI] 1.29-1.67, P = .032), benefit from revascularization clinical stage 3 on SVS-WIfI score (RR 4.56: 95%CI 1.21-17.21, P = 0.007), and high WM classification score (RR 9.46: 95% CI 5.65-15.82, P < 0.001) were associated by multivariate analysis with high amputation rates. 1-year & 3-year amputation risk on SVS-WIfI score were not associated with amputation rates in DFI (P = .263 and .496). Only 9 (6.8%) patients were lost to follow up during the 3-year period. WM classification score, SVS-WIfI score on benefit from revascularization, and PAD were strongly associated with major amputation rates in patients with DFI.

A characteristic cerebellar biosignature for bipolar disorder, identified with fully automatic machine learning

BackgroundTranscriptomic profile differences between patients with bipolar disorder and healthy controls can be identified using machine learning and can provide information about the potential role of the cerebellum in the pathogenesis of bipolar disorder.With this aim, user-friendly, fully automated machine learning algorithms can achieve extremely high classification scores and disease-related predictive biosignature identification, in short time frames and scaled down to small datasets. MethodA fully automated machine learning platform, based on the most suitable algorithm selection and relevant set of hyper-parameter values, was applied on a preprocessed transcriptomics dataset, in order to produce a model for biosignature selection and to classify subjects into groups of patients and controls. The parent GEO datasets were originally produced from the cerebellar and parietal lobe tissue of deceased bipolar patients and healthy controls, using Affymetrix Human Gene 1.0 ST Array. ResultsPatients and controls were classified into two separate groups, with no close-to-the-boundary cases, and this classification was based on the cerebellar transcriptomic biosignature of 25 features (genes), with Area Under Curve 0.929 and Average Precision 0.955. The biosignature includes both genes connected before to bipolar disorder, depression, psychosis or epilepsy, as well as genes not linked before with any psychiatric disease. Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis revealed participation of 4 identified features in 6 pathways which have also been associated with bipolar disorder. ConclusionAutomated machine learning (AutoML) managed to identify accurately 25 genes that can jointly – in a multivariate–fashion - separate bipolar patients from healthy controls with high predictive power. The discovered features lead to new biological insights. Machine Learning (ML) analysis considers the features in combination (in contrast to standard differential expression analysis), removing both irrelevant as well as redundant markers, and thus, focusing to biological interpretation.

Robust object tracking via ensembling semantic‐aware network and redetection

AbstractMost Siamese‐based trackers use classification and regression to determine the target bounding box, which can be formulated as a linear matching process of the template and search region. However, this only takes into account the similarity of features while ignoring the semantic object information, resulting in some cases in which the regression box with the highest classification score is not accurate. To address the lack of semantic information, an object tracking approach based on an ensemble semantic‐aware network and redetection (ESART) is proposed. Furthermore, a DarkNet53 network with transfer learning is used as our semantic‐aware model to adapt the detection task for extracting semantic information. In addition, a semantic tag redetection method to re‐evaluate the bounding box and overcome inaccurate scaling issues is proposed. Extensive experiments based on OTB2015, UAV123, UAV20L, and GOT‐10k show that our tracker is superior to other state‐of‐the‐art trackers. It is noteworthy that our semantic‐aware ensemble method can be embedded into any tracker for classification and regression task.

Benchmarking framework for machine learning classification from fNIRS data.

While efforts to establish best practices with functional near infrared spectroscopy (fNIRS) signal processing have been published, there are still no community standards for applying machine learning to fNIRS data. Moreover, the lack of open source benchmarks and standard expectations for reporting means that published works often claim high generalisation capabilities, but with poor practices or missing details in the paper. These issues make it hard to evaluate the performance of models when it comes to choosing them for brain-computer interfaces. We present an open-source benchmarking framework, BenchNIRS, to establish a best practice machine learning methodology to evaluate models applied to fNIRS data, using five open access datasets for brain-computer interface (BCI) applications. The BenchNIRS framework, using a robust methodology with nested cross-validation, enables researchers to optimise models and evaluate them without bias. The framework also enables us to produce useful metrics and figures to detail the performance of new models for comparison. To demonstrate the utility of the framework, we present a benchmarking of six baseline models [linear discriminant analysis (LDA), support-vector machine (SVM), k-nearest neighbours (kNN), artificial neural network (ANN), convolutional neural network (CNN), and long short-term memory (LSTM)] on the five datasets and investigate the influence of different factors on the classification performance, including: number of training examples and size of the time window of each fNIRS sample used for classification. We also present results with a sliding window as opposed to simple classification of epochs, and with a personalised approach (within subject data classification) as opposed to a generalised approach (unseen subject data classification). Results show that the performance is typically lower than the scores often reported in literature, and without great differences between models, highlighting that predicting unseen data remains a difficult task. Our benchmarking framework provides future authors, who are achieving significant high classification scores, with a tool to demonstrate the advances in a comparable way. To complement our framework, we contribute a set of recommendations for methodology decisions and writing papers, when applying machine learning to fNIRS data.

Automated localization of the medial clavicular epiphyseal cartilages using an object detection network: a step towards deep learning-based forensic age assessment

BackgroundDeep learning is a promising technique to improve radiological age assessment. However, expensive manual annotation by experts poses a bottleneck for creating large datasets to appropriately train deep neural networks. We propose an object detection approach to automatically annotate the medial clavicular epiphyseal cartilages in computed tomography (CT) scans.MethodsThe sternoclavicular joints were selected as structure-of-interest (SOI) in chest CT scans and served as an easy-to-identify proxy for the actual medial clavicular epiphyseal cartilages. CT slices containing the SOI were manually annotated with bounding boxes around the SOI. All slices in the training set were used to train the object detection network RetinaNet. Afterwards, the network was applied individually to all slices of the test scans for SOI detection. Bounding box and slice position of the detection with the highest classification score were used as the location estimate for the medial clavicular epiphyseal cartilages inside the CT scan.ResultsFrom 100 CT scans of 82 patients, 29,656 slices were used for training and 30,846 slices from 110 CT scans of 110 different patients for testing the object detection network. The location estimate from the deep learning approach for the SOI was in a correct slice in 97/110 (88%), misplaced by one slice in 5/110 (5%), and missing in 8/110 (7%) test scans. No estimate was misplaced by more than one slice.ConclusionsWe demonstrated a robust automated approach for annotating the medial clavicular epiphyseal cartilages. This enables training and testing of deep neural networks for age assessment.

An IoU-aware Siamese network for real-time visual tracking

Most Siamese trackers decompose the tracking task into two branches: classification and bounding box regression. They choose the bounding box with the highest classification score or the combination of the classification and predicted localization scores as the target. However, the misalignment between the classification and localization accuracy will degrade tracking performance. In this paper, we propose an IoU-aware Siamese tracker named IASNet, which predicts the IoU classification score of each regressed box to represent its localization confidence. We introduce a novel residual alignment module to capture the scale of each predicted bounding box and its border context, which further improves the reliability of the classification prediction. In addition, we build dynamic links between the classification and regression branches to enhance their interaction. The proposed dynamic loss enables the training to focus on high-quality positive samples, leading to more precise localization. Extensive experimental results on VOT2016, VOT2019, OTB100, GOT10k, UAV123, and LaSOT show that the proposed IASNet achieves state-of-the-art tracking performance and runs at approximately 65 fps on a GTX 3090 GPU, which far exceeds the real-time requirement.

Classification of Kinematic and Electromyographic Signals Associated with Pathological Tremor Using Machine and Deep Learning.

Peripheral Electrical Stimulation (PES) of afferent pathways has received increased interest as a solution to reduce pathological tremors with minimal side effects. Closed-loop PES systems might present some advantages in reducing tremors, but further developments are required in order to reliably detect pathological tremors to accurately enable the stimulation only if a tremor is present. This study explores different machine learning (K-Nearest Neighbors, Random Forest and Support Vector Machines) and deep learning (Long Short-Term Memory neural networks) models in order to provide a binary (Tremor; No Tremor) classification of kinematic (angle displacement) and electromyography (EMG) signals recorded from patients diagnosed with essential tremors and healthy subjects. Three types of signal sequences without any feature extraction were used as inputs for the classifiers: kinematics (wrist flexion-extension angle), raw EMG and EMG envelopes from wrist flexor and extensor muscles. All the models showed high classification scores (Tremor vs. No Tremor) for the different input data modalities, ranging from 0.8 to 0.99 for the f1 score. The LSTM models achieved 0.98 f1 scores for the classification of raw EMG signals, showing high potential to detect tremors without any processed features or preliminary information. These models may be explored in real-time closed-loop PES strategies to detect tremors and enable stimulation with minimal signal processing steps.

A Framework to Rank Prognostics Health Indicators with Application to Brake Rotors

This study presents a framework to assess the effectiveness of various health indicators (HIs) used to monitor the state of health (SOH) of a brake rotor health monitoring system. The following criteria were used to rank various health indicators: (i) Identifiability: Correlation of the HI with the Ground Truth (GT); (ii) Compactness: Mean of the standard deviation of the estimated SOHs; (iii) Robustness to Noise Factors: An HI is considered robust when it meets all functional and customer requirements under all operating conditions and its performance is not affected by the variations in the environment, operating conditions or other factors impacting the performance in an undesired way (noise factors); (iv) Monotonicity: To quantify the monotonic trend in HIs as the fault level increases from healthy baseline to the most severe faults. Monotone HIs are preferred as they will likely generalize better to data not used in development; and (v) Estimation Error: The average relative error between the GT and the prediction obtained from the regression analysis. Results showed that this framework can be applied to several HIs derived from performing time and frequency analysis on various sensor signals used to monitor the health of brake rotors. Top HIs selected based on this framework provided the best performance in detecting degraded brake rotors as evidenced by higher classification score.

DeepNoise: Signal and Noise Disentanglement Based on Classifying Fluorescent Microscopy Images via Deep Learning

The high-content image-based assay is commonly leveraged for identifying the phenotypic impact of genetic perturbations in biology field. However, a persistent issue remains unsolved during experiments: the interferential technical noises caused by systematic errors (e.g., temperature, reagent concentration, and well location) are always mixed up with the real biological signals, leading to misinterpretation of any conclusion drawn. Here, we reported a mean teacher-based deep learning model (DeepNoise) that can disentangle biological signals from the experimental noises. Specifically, we aimed to classify the phenotypic impact of 1108 different genetic perturbations screened from 125,510 fluorescent microscopy images, which were totally unrecognizable by the human eye. We validated our model by participating in the Recursion Cellular Image Classification Challenge, and DeepNoise achieved an extremely high classification score (accuracy: 99.596%), ranking the 2nd place among 866 participating groups. This promising result indicates the successful separation of biological and technical factors, which might help decrease the cost of treatment development and expedite the drug discovery process. The source code of DeepNoise is available at https://github.com/Scu-sen/Recursion-Cellular-Image-Classification-Challenge.

Artificial Neural Network-Based Caprock Structural Reliability Analysis for CO2 Injection Site—An Example from Northern North Sea

In CO2 sequestration projects, assessing caprock structural stability is crucial to assure the success and reliability of the CO2 injection. However, since caprock experimental data are sparse, we applied a Monte Carlo (MC) algorithm to generate stochastic data from the given mean and standard deviation values. The generated data sets were introduced to a neural network (NN), including four hidden layers for classification purposes. The model was then used to evaluate organic-rich Draupne caprock shale failure in the Alpha structure, northern North Sea. The train and test were carried out with 75% and 25% of the input data, respectively. Following that, validation is accomplished with unseen data, yielding promising classification scores. The results show that introducing larger input data sizes to the established NN provides better convergence conditions and higher classification scores. Although the NN can predicts the failure states with a classification score of 97%, the structural reliability was significantly low compare to the failure results estimated using other method. Moreover, this indicated that during evaluating the field-scale caprock failure, more experimental data is needed for a reliable result. However, this study depicts the advantage of machine learning algorithms in geological CO2 storage projects compared with similar finite elements methods in the aspect of short fitting time, high accuracy, and flexibility in processing different input data sizes with different scales.

IoU-Balanced loss functions for single-stage object detection

Single-stage object detectors have been widely applied in computer vision applications due to their high efficiency. However, the loss functions adopted by single-stage detectors hurt the localization accuracy seriously. Firstly, the cross-entropy loss for classification is independent of the localization task and drives all the positive examples to learn as high classification scores as possible regardless of localization accuracy. Thus, there exist many detections with high classification scores but low IoU or detections with low classification scores but high IoU. Secondly, for the smooth L1 loss, the gradient is dominated by the outliers with poor localization accuracy. In this work, IoU-balanced loss functions consisting of IoU-balanced classification loss and IoU-balanced localization loss are proposed to solve these problems. IoU-balanced classification loss pays more attention to positive examples with high IoU and enhances the correlation between classification and localization tasks. IoU-balanced localization loss decreases the gradient of examples with low IoU and increases the gradient of examples with high IoU, which improves the localization accuracy of models. Extensive experiments on MS COCO, PASCAL VOC, Cityscapes and WIDERFace demonstrate that IoU-balanced losses can substantially improve the popular single-stage detectors, especially the localization accuracy. On COCO test-dev , the proposed methods can substantially improve AP by 1.0 % ∼ 1.7 % and AP 75 by 1.0 % ∼ 2.4 % . On PASCAL VOC, Cityscape and WIDERFace, it can also substantially improve AP by 1.0 % ∼ 1.5 % and A P 80 , A P 90 by ∼ 3.9 % . The source code will be made publicly available.

Machine Learning Meets Compressed Sensing in Vibration-Based Monitoring.

Artificial Intelligence applied to Structural Health Monitoring (SHM) has provided considerable advantages in the accuracy and quality of the estimated structural integrity. Nevertheless, several challenges still need to be tackled in the SHM field, which extended the monitoring process beyond the mere data analytics and structural assessment task. Besides, one of the open problems in the field relates to the communication layer of the sensor networks since the continuous collection of long time series from multiple sensing units rapidly consumes the available memory resources, and requires complicated protocol to avoid network congestion. In this scenario, the present work presents a comprehensive framework for vibration-based diagnostics, in which data compression techniques are firstly introduced as a means to shrink the dimension of the data to be managed through the system. Then, neural network models solving binary classification problems were implemented for the sake of damage detection, also encompassing the influence of environmental factors in the evaluation of the structural status. Moreover, the potential degradation induced by the usage of low cost sensors on the adopted framework was evaluated: Additional analyses were performed in which experimental data were corrupted with the noise characterizing MEMS sensors. The proposed solutions were tested with experimental data from the Z24 bridge use case, proving that the amalgam of data compression, optimized (i.e., low complexity) machine learning architectures and environmental information allows to attain high classification scores, i.e., accuracy and precision greater than 96% and 95%, respectively.

Influence of Image Enhancement Techniques on Effectiveness of Unconstrained Face Detection and Identification

In a criminal investigation, along with processing forensic evidence, different investigative techniques are used to identify the perpetrator of the crime. It includes collecting and analyzing unconstrained face images, mostly with low resolution and various qualities, making identification difficult. Since police organizations have limited resources, in this paper, we propose a novel method that utilizes off-the-shelf solutions (Dlib library Histogram of Oriented Gradients-HOG face detectors and the ResNet faces feature vector extractor) to provide practical assistance in unconstrained face identification. Our experiment aimed to establish which one (if any) of the basic image enhancement techniques should be applied to increase the effectiveness. Results obtained from three publicly available databases and one created for this research (simulating police investigators’ database) showed that resizing the image (especially with a resolution lower than 150 pixels) should always precede enhancement to improve face detection accuracy. The best results in determining whether they are the same or different persons in images were obtained by applying sharpening with a high-pass filter, whereas normalization gives the highest classification scores when a single weight value is applied to data from all four databases.

Towards Model Generalization for Intrusion Detection: Unsupervised Machine Learning Techniques

Through the ongoing digitization of the world, the number of connected devices is continuously growing without any foreseen decline in the near future. In particular, these devices increasingly include critical systems such as power grids and medical institutions, possibly causing tremendous consequences in the case of a successful cybersecurity attack. A network intrusion detection system (NIDS) is one of the main components to detect ongoing attacks by differentiating normal from malicious traffic. Anomaly-based NIDS, more specifically unsupervised methods previously proved promising for their ability to detect known as well as zero-day attacks without the need for a labeled dataset. Despite decades of development by researchers, anomaly-based NIDS are only rarely employed in real-world applications, most possibly due to the lack of generalization power of the proposed models. This article first evaluates four unsupervised machine learning methods on two recent datasets and then defines their generalization strength using a novel inter-dataset evaluation strategy estimating their adaptability. Results show that all models can present high classification scores on an individual dataset but fail to directly transfer those to a second unseen but related dataset. Specifically, the accuracy dropped on average 25.63% in an inter-dataset setting compared to the conventional evaluation approach. This generalization challenge can be observed and tackled in future research with the help of the proposed evaluation strategy in this paper.

Automatic detection and classification of baleen whale social calls using convolutional neural networks.

Passive acoustic monitoring has proven to be an indispensable tool for many aspects of baleen whale research. Manual detection of whale calls on these large data sets demands extensive manual labor. Automated whale call detectors offer a more efficient approach and have been developed for many species and call types. However, calls with a large level of variability such as fin whale (Balaenoptera physalus) 40 Hz call and blue whale (B. musculus) D call have been challenging to detect automatically and hence no practical automated detector exists for these two call types. Using a modular approach consisting of faster region-based convolutional neural network followed by a convolutional neural network, we have created automated detectors for 40 Hz calls and D calls. Both detectors were tested on recordings with high- and low density of calls and, when selecting for detections with high classification scores, they were shown to have precision ranging from 54% to 57% with recall ranging from 72% to 78% for 40 Hz and precision ranging from 62% to 64% with recall ranging from 70 to 73% for D calls. As these two call types are produced by both sexes, using them in long-term studies would remove sex-bias in estimates of temporal presence and movement patterns.

From the Laboratory to the Field: IMU-Based Shot and Pass Detection in Football Training and Game Scenarios Using Deep Learning.

The applicability of sensor-based human activity recognition in sports has been repeatedly shown for laboratory settings. However, the transferability to real-world scenarios cannot be granted due to limitations on data and evaluation methods. On the example of football shot and pass detection against a null class we explore the influence of those factors for real-world event classification in field sports. For this purpose we compare the performance of an established Support Vector Machine (SVM) for laboratory settings from literature to the performance in three evaluation scenarios gradually evolving from laboratory settings to real-world scenarios. In addition, three different types of neural networks, namely a convolutional neural net (CNN), a long short term memory net (LSTM) and a convolutional LSTM (convLSTM) are compared. Results indicate that the SVM is not able to reliably solve the investigated three-class problem. In contrast, all deep learning models reach high classification scores showing the general feasibility of event detection in real-world sports scenarios using deep learning. The maximum performance with a weighted f1-score of 0.93 was reported by the CNN. The study provides valuable insights for sports assessment under practically relevant conditions. In particular, it shows that (1) the discriminative power of established features needs to be reevaluated when real-world conditions are assessed, (2) the selection of an appropriate dataset and evaluation method are both required to evaluate real-world applicability and (3) deep learning-based methods yield promising results for real-world HAR in sports despite high variations in the execution of activities.