Pub Date : 2021-11-01DOI: 10.23919/fusion49465.2021.9626979
Sehyun Yun, Renato Zanetti
A nonlinear filter based on an artificial neural network (ANN) is proposed to accurately estimate the state of a nonlinear dynamic system. The ANN is trained to learn the nonlinear mapping between the inputs and outputs of training data. The proposed filter is computationally efficient for online applications because estimation error can be directly estimated once the ANN is trained offline. The unscented transformation (UT) is employed in this filter to approximate the first two moments of the estimate. Under the scenarios considered in this paper, it is shown through numerical simulation that the proposed filter leads to better performance than the extended Kalman filter (EKF), unscented Kalman filter (UKF), and a state-of-the-art nonlinear filter.
{"title":"Bayesian Estimation with Artificial Neural Network","authors":"Sehyun Yun, Renato Zanetti","doi":"10.23919/fusion49465.2021.9626979","DOIUrl":"https://doi.org/10.23919/fusion49465.2021.9626979","url":null,"abstract":"A nonlinear filter based on an artificial neural network (ANN) is proposed to accurately estimate the state of a nonlinear dynamic system. The ANN is trained to learn the nonlinear mapping between the inputs and outputs of training data. The proposed filter is computationally efficient for online applications because estimation error can be directly estimated once the ANN is trained offline. The unscented transformation (UT) is employed in this filter to approximate the first two moments of the estimate. Under the scenarios considered in this paper, it is shown through numerical simulation that the proposed filter leads to better performance than the extended Kalman filter (EKF), unscented Kalman filter (UKF), and a state-of-the-art nonlinear filter.","PeriodicalId":226850,"journal":{"name":"2021 IEEE 24th International Conference on Information Fusion (FUSION)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128809802","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-11-01DOI: 10.23919/fusion49465.2021.9627045
C. Chong
Traditional multiple target tracking (MTT) algorithms are model-based. Target and sensor models are used to associate measurements, perform track filtering, score possible associations, and find the best association hypothesis. Recent advances in machine learning (ML) have resulted in data-driven model-free methods for MTT, especially in computer vision, where MTT is called multiple object tracking (MOT). This paper presents an overview of ML methods for detection, track filtering, data association, and end-to-end MTT. It assesses the state-of-the-art and presents future research directions.
{"title":"An Overview of Machine Learning Methods for Multiple Target Tracking","authors":"C. Chong","doi":"10.23919/fusion49465.2021.9627045","DOIUrl":"https://doi.org/10.23919/fusion49465.2021.9627045","url":null,"abstract":"Traditional multiple target tracking (MTT) algorithms are model-based. Target and sensor models are used to associate measurements, perform track filtering, score possible associations, and find the best association hypothesis. Recent advances in machine learning (ML) have resulted in data-driven model-free methods for MTT, especially in computer vision, where MTT is called multiple object tracking (MOT). This paper presents an overview of ML methods for detection, track filtering, data association, and end-to-end MTT. It assesses the state-of-the-art and presents future research directions.","PeriodicalId":226850,"journal":{"name":"2021 IEEE 24th International Conference on Information Fusion (FUSION)","volume":"11 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131404210","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-11-01DOI: 10.23919/FUSION49465.2021.9627024
André Brandenburger, Folker Hoffmann, A. Charlish
Reinforcement learning (RL) is already widely applied to applications such as robotics, but it is only sparsely used in sensor management. In this paper, we apply the popular Proximal Policy Optimization (PPO) approach to a multi-agent UAV tracking scenario. While recorded data of real scenarios can accurately reflect the real world, the required amount of data is not always available. Simulation data, however, is typically cheap to generate, but the utilized target behavior is often naive and only vaguely represents the real world. In this paper, we utilize multi-agent RL to jointly generate protagonistic and antagonistic policies and overcome the data generation problem, as the policies are generated on-the-fly and adapt continuously. This way, we are able to clearly outperform baseline methods and robustly generate competitive policies. In addition, we investigate explainable artificial intelligence (XAI) by interpreting feature saliency and generating an easy-to-read decision tree as a simplified policy.
{"title":"Co-Training an Observer and an Evading Target","authors":"André Brandenburger, Folker Hoffmann, A. Charlish","doi":"10.23919/FUSION49465.2021.9627024","DOIUrl":"https://doi.org/10.23919/FUSION49465.2021.9627024","url":null,"abstract":"Reinforcement learning (RL) is already widely applied to applications such as robotics, but it is only sparsely used in sensor management. In this paper, we apply the popular Proximal Policy Optimization (PPO) approach to a multi-agent UAV tracking scenario. While recorded data of real scenarios can accurately reflect the real world, the required amount of data is not always available. Simulation data, however, is typically cheap to generate, but the utilized target behavior is often naive and only vaguely represents the real world. In this paper, we utilize multi-agent RL to jointly generate protagonistic and antagonistic policies and overcome the data generation problem, as the policies are generated on-the-fly and adapt continuously. This way, we are able to clearly outperform baseline methods and robustly generate competitive policies. In addition, we investigate explainable artificial intelligence (XAI) by interpreting feature saliency and generating an easy-to-read decision tree as a simplified policy.","PeriodicalId":226850,"journal":{"name":"2021 IEEE 24th International Conference on Information Fusion (FUSION)","volume":"266 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115599085","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-11-01DOI: 10.23919/fusion49465.2021.9626972
Xueqian Wang, D. Zhu, Gang Li, Xiao-Ping Zhang
In this paper, we investigate the fusion of spaceborne synthetic aperture radar (SAR) and airborne SAR images and its application to ship target enhancement. In this paper, we propose a new target proposal and clutter copula (TPCC)-based image fusion method for the collaboration of spaceborne and airborne SARs. TPCC enhances the common ship target areas in spaceborne and airborne SAR images via the intersection of target proposals and suppresses the clutter areas by establishing the joint distribution of clutter in the spaceborne and airborne SAR images based on the copula theory. Compared with other commonly used image fusion methods, the target dependence and clutter dependence in the spaceborne and airborne SAR images are newly exploited in TPCC. We demonstrate the superiority of TPCC in terms of target-to-clutter ratios (TCRs) by using composite images combining Gaofen-3 satellite and unmanned aerial vehicle (UAV) SAR images.
{"title":"A New Image Fusion Method for Ship Target Enhancement in Spaceborne and Airborne SAR Collaboration","authors":"Xueqian Wang, D. Zhu, Gang Li, Xiao-Ping Zhang","doi":"10.23919/fusion49465.2021.9626972","DOIUrl":"https://doi.org/10.23919/fusion49465.2021.9626972","url":null,"abstract":"In this paper, we investigate the fusion of spaceborne synthetic aperture radar (SAR) and airborne SAR images and its application to ship target enhancement. In this paper, we propose a new target proposal and clutter copula (TPCC)-based image fusion method for the collaboration of spaceborne and airborne SARs. TPCC enhances the common ship target areas in spaceborne and airborne SAR images via the intersection of target proposals and suppresses the clutter areas by establishing the joint distribution of clutter in the spaceborne and airborne SAR images based on the copula theory. Compared with other commonly used image fusion methods, the target dependence and clutter dependence in the spaceborne and airborne SAR images are newly exploited in TPCC. We demonstrate the superiority of TPCC in terms of target-to-clutter ratios (TCRs) by using composite images combining Gaofen-3 satellite and unmanned aerial vehicle (UAV) SAR images.","PeriodicalId":226850,"journal":{"name":"2021 IEEE 24th International Conference on Information Fusion (FUSION)","volume":"212 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115776567","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-11-01DOI: 10.23919/fusion49465.2021.9626955
Luisa Still, M. Oispuu, W. Koch
This paper proposes a method to find an optimal set of sensor positions for the shooter localization task. Here, optimality is defined in terms of best possible state estimation accuracy given by the Cramér-Rao bound. We derive an optimality criterion, present an application specific genetic algorithm to solve the optimization problem and investigate different scenarios with complete and incomplete measurement data sets and varying number of sensors. As an intermediate step we assume that the shooter state is exactly known. The results show that depending on the available measurement data set, the recommended optimal sensor positions are often unexpected. For all considered scenarios, the applied optimization approach determines the optimal positions reliably.
{"title":"Optimal Sensor Placement for Shooter Localization Using a Genetic Algorithm","authors":"Luisa Still, M. Oispuu, W. Koch","doi":"10.23919/fusion49465.2021.9626955","DOIUrl":"https://doi.org/10.23919/fusion49465.2021.9626955","url":null,"abstract":"This paper proposes a method to find an optimal set of sensor positions for the shooter localization task. Here, optimality is defined in terms of best possible state estimation accuracy given by the Cramér-Rao bound. We derive an optimality criterion, present an application specific genetic algorithm to solve the optimization problem and investigate different scenarios with complete and incomplete measurement data sets and varying number of sensors. As an intermediate step we assume that the shooter state is exactly known. The results show that depending on the available measurement data set, the recommended optimal sensor positions are often unexpected. For all considered scenarios, the applied optimization approach determines the optimal positions reliably.","PeriodicalId":226850,"journal":{"name":"2021 IEEE 24th International Conference on Information Fusion (FUSION)","volume":"39 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"124365876","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-11-01DOI: 10.23919/fusion49465.2021.9627066
N.D. Blomerus, J. D. Villiers, Willie Nel
There have been numerous advancements in machine learning technologies in recent years, which has led to the application of machine learning algorithms to automatic target recognition. Two key challenges for these methods are the lack of sufficient training datasets and non-transparent deep models. In this paper, experiments are conducted that investigate the application of target detection using a model trained on the MSTAR to detect targets in another dataset, as well as the investigation of uncertainty estimates in Bayesian convolutional neural networks and how these outputs can improve confidence in the model’s predictions. The model can correctly detect targets in the test scene’s, as well as targets not seen from the MSTAR dataset. The output of the Bayesian convolutional neural network is used to create uncertainty heat maps. The epistemic uncertainty is the uncertainty created by the model and aleatoric is created by the data. These heat maps are overlaid on SAR images, thereby aiding in explainability by highlighting regions in the SAR images that exhibit high uncertainty from a classification point of view. Hence, uncertainty estimates from the Bayesian model give insight into the confidence of its predictions and show promise to improve trust between users and the model.
{"title":"Improved Explainability through Uncertainty Estimation in Automatic Target Recognition of SAR Images","authors":"N.D. Blomerus, J. D. Villiers, Willie Nel","doi":"10.23919/fusion49465.2021.9627066","DOIUrl":"https://doi.org/10.23919/fusion49465.2021.9627066","url":null,"abstract":"There have been numerous advancements in machine learning technologies in recent years, which has led to the application of machine learning algorithms to automatic target recognition. Two key challenges for these methods are the lack of sufficient training datasets and non-transparent deep models. In this paper, experiments are conducted that investigate the application of target detection using a model trained on the MSTAR to detect targets in another dataset, as well as the investigation of uncertainty estimates in Bayesian convolutional neural networks and how these outputs can improve confidence in the model’s predictions. The model can correctly detect targets in the test scene’s, as well as targets not seen from the MSTAR dataset. The output of the Bayesian convolutional neural network is used to create uncertainty heat maps. The epistemic uncertainty is the uncertainty created by the model and aleatoric is created by the data. These heat maps are overlaid on SAR images, thereby aiding in explainability by highlighting regions in the SAR images that exhibit high uncertainty from a classification point of view. Hence, uncertainty estimates from the Bayesian model give insight into the confidence of its predictions and show promise to improve trust between users and the model.","PeriodicalId":226850,"journal":{"name":"2021 IEEE 24th International Conference on Information Fusion (FUSION)","volume":"4 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121968734","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-11-01DOI: 10.23919/fusion49465.2021.9626895
Tobias Fleck, Johann Marius Zöllner
Tracking-by-Detection has become the major paradigm in Multi Object Tracking (MOT) for a large variety of sensors. Regardless of the type of tracking system, hyper parameters are often chosen manually instead of doing a structured search to reveal the full potential of the system.In this work we tackle this problem by utilizing Bayesian Optimization (BO) to tune tracking systems, enabling to find the best combination of hyper parameters for Gaussian Mixture Probability Hypothesis Density Trackers (GM-PHD) in two different tracking applications. We use the Tree-structured Parzen Estimator (TPE) algorithm [1] [2] with an Expected Improvement (EI) acquisition function as a blackbox optimizer. TPE supports to conveniently incorporate domain expert knowledge by modeling prior probability distributions of the search space. In our experiments we use the popular MOTA metric as optimization objective.Evaluation is performed in a simulation scenario with an in depth discussion of the found parameters and a real world example that uses the MOT-20 challenge dataset [3] demonstrates the unconditional applicability of the approach. We finish with a conclusion on Bayesian Optimization for MOT systems and future research.
{"title":"Tuning Multi Object Tracking Systems using Bayesian Optimization","authors":"Tobias Fleck, Johann Marius Zöllner","doi":"10.23919/fusion49465.2021.9626895","DOIUrl":"https://doi.org/10.23919/fusion49465.2021.9626895","url":null,"abstract":"Tracking-by-Detection has become the major paradigm in Multi Object Tracking (MOT) for a large variety of sensors. Regardless of the type of tracking system, hyper parameters are often chosen manually instead of doing a structured search to reveal the full potential of the system.In this work we tackle this problem by utilizing Bayesian Optimization (BO) to tune tracking systems, enabling to find the best combination of hyper parameters for Gaussian Mixture Probability Hypothesis Density Trackers (GM-PHD) in two different tracking applications. We use the Tree-structured Parzen Estimator (TPE) algorithm [1] [2] with an Expected Improvement (EI) acquisition function as a blackbox optimizer. TPE supports to conveniently incorporate domain expert knowledge by modeling prior probability distributions of the search space. In our experiments we use the popular MOTA metric as optimization objective.Evaluation is performed in a simulation scenario with an in depth discussion of the found parameters and a real world example that uses the MOT-20 challenge dataset [3] demonstrates the unconditional applicability of the approach. We finish with a conclusion on Bayesian Optimization for MOT systems and future research.","PeriodicalId":226850,"journal":{"name":"2021 IEEE 24th International Conference on Information Fusion (FUSION)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128696248","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-11-01DOI: 10.23919/fusion49465.2021.9626896
Nicholas Pym, A. D. Freitas
Systems capable of intelligently monitoring the traffic of people at entrances to enclosed areas enable a variety of useful applications such as improved retail store analytics. However, the real-world implementation of such a system is typically hindered by computationally expensive algorithms and privacy concerns. In this paper, a low-cost privacy-sensitive intelligent monitoring system based on an embedded platform is presented. The key components of the system include a people classification model and a people re-identification model. A detailed description of the optimisation of these components is presented. The developed system is able to detect people entering/exiting a closed area with an accuracy above 99% in real-time. In addition, the system is able to achieve re-identification accuracy above 93% in under 0.7 seconds on an embedded system. Data collected by the system was used for training and it was tested under real-world conditions.
{"title":"An embedded platform approach to privacy-centric person re-identification","authors":"Nicholas Pym, A. D. Freitas","doi":"10.23919/fusion49465.2021.9626896","DOIUrl":"https://doi.org/10.23919/fusion49465.2021.9626896","url":null,"abstract":"Systems capable of intelligently monitoring the traffic of people at entrances to enclosed areas enable a variety of useful applications such as improved retail store analytics. However, the real-world implementation of such a system is typically hindered by computationally expensive algorithms and privacy concerns. In this paper, a low-cost privacy-sensitive intelligent monitoring system based on an embedded platform is presented. The key components of the system include a people classification model and a people re-identification model. A detailed description of the optimisation of these components is presented. The developed system is able to detect people entering/exiting a closed area with an accuracy above 99% in real-time. In addition, the system is able to achieve re-identification accuracy above 93% in under 0.7 seconds on an embedded system. Data collected by the system was used for training and it was tested under real-world conditions.","PeriodicalId":226850,"journal":{"name":"2021 IEEE 24th International Conference on Information Fusion (FUSION)","volume":"57 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130580749","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
International maritime crime is becoming increasingly sophisticated, often associated with wider criminal networks. Detecting maritime threats by means of fusing data purely related to physical movement (i.e., those generated by physical sensors, or hard data) is not sufficient. This has led to research and development efforts aimed at combining hard data with other types of data (especially human-generated or soft data). Existing work often assumes that input soft data is available in a structured format, or is focused on extracting certain relevant entities or concepts to accompany or annotate hard data. Much less attention has been given to extracting the rich knowledge about the situations of interest implicitly embedded in the large amount of soft data existing in unstructured formats (such as intelligence reports and news articles). In order to exploit the potentially useful and rich information from such sources, it is necessary to extract not only the relevant entities and concepts, but also their semantic relations, together with the uncertainty associated with the extracted knowledge (i.e., in the form of probabilistic knowledge graphs). This will increase the accuracy of, and confidence in, the extracted knowledge and facilitate subsequent reasoning and learning. To this end, we propose Maritime DeepDive, an initial prototype for the automated construction of probabilistic knowledge graphs from natural language data for the maritime domain. In this paper, we report on the current implementation of Maritime DeepDive, together with preliminary results on extracting probabilistic events from maritime piracy incidents. This pipeline was evaluated on a manually crafted gold standard, yielding promising results.
{"title":"Toward the Automated Construction of Probabilistic Knowledge Graphs for the Maritime Domain","authors":"Fatemeh Shiri, Teresa Wang, Shirui Pan, Xiaojun Chang, Yuan-Fang Li, Gholamreza Haffari, V. Nguyen, Shuang Yu","doi":"10.23919/FUSION49465.2021.9626935","DOIUrl":"https://doi.org/10.23919/FUSION49465.2021.9626935","url":null,"abstract":"International maritime crime is becoming increasingly sophisticated, often associated with wider criminal networks. Detecting maritime threats by means of fusing data purely related to physical movement (i.e., those generated by physical sensors, or hard data) is not sufficient. This has led to research and development efforts aimed at combining hard data with other types of data (especially human-generated or soft data). Existing work often assumes that input soft data is available in a structured format, or is focused on extracting certain relevant entities or concepts to accompany or annotate hard data. Much less attention has been given to extracting the rich knowledge about the situations of interest implicitly embedded in the large amount of soft data existing in unstructured formats (such as intelligence reports and news articles). In order to exploit the potentially useful and rich information from such sources, it is necessary to extract not only the relevant entities and concepts, but also their semantic relations, together with the uncertainty associated with the extracted knowledge (i.e., in the form of probabilistic knowledge graphs). This will increase the accuracy of, and confidence in, the extracted knowledge and facilitate subsequent reasoning and learning. To this end, we propose Maritime DeepDive, an initial prototype for the automated construction of probabilistic knowledge graphs from natural language data for the maritime domain. In this paper, we report on the current implementation of Maritime DeepDive, together with preliminary results on extracting probabilistic events from maritime piracy incidents. This pipeline was evaluated on a manually crafted gold standard, yielding promising results.","PeriodicalId":226850,"journal":{"name":"2021 IEEE 24th International Conference on Information Fusion (FUSION)","volume":"31 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"132508589","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-11-01DOI: 10.23919/fusion49465.2021.9627071
Shenglin Wang, Peng Wang, L. Mihaylova, Matthew Hill
Deep neural networks (DNNs) have become very popular recently and have proven their potential especially for image classification. However, their performance depends significantly on the network structure and data quality. This paper investigates the performance of DNNs and especially of faster region based convolutional neural networks (R-CNNs), called faster R-CNN when the network testing data differ significantly from the training data. This paper proposes a framework for monitoring the neuron patterns within a faster R-CNN by representing distributions of neuron activation patterns and by calculating corresponding distances between them, with the Kullback-Leibler divergence. The patterns of the activation states of ‘neurons’ within the network can therefore be observed if the faster R-CNN is ‘outside the comfort zone’, mostly when it works with noisy data and data that are significantly different from those used in the training stage. The validation is performed on publicly available datasets: MNIST [1] and PASCAL [2] and demonstrates that the proposed framework can be used for real-time monitoring of supervised classifiers.
{"title":"Real-time Activation Pattern Monitoring and Uncertainty Characterisation in Image Classification","authors":"Shenglin Wang, Peng Wang, L. Mihaylova, Matthew Hill","doi":"10.23919/fusion49465.2021.9627071","DOIUrl":"https://doi.org/10.23919/fusion49465.2021.9627071","url":null,"abstract":"Deep neural networks (DNNs) have become very popular recently and have proven their potential especially for image classification. However, their performance depends significantly on the network structure and data quality. This paper investigates the performance of DNNs and especially of faster region based convolutional neural networks (R-CNNs), called faster R-CNN when the network testing data differ significantly from the training data. This paper proposes a framework for monitoring the neuron patterns within a faster R-CNN by representing distributions of neuron activation patterns and by calculating corresponding distances between them, with the Kullback-Leibler divergence. The patterns of the activation states of ‘neurons’ within the network can therefore be observed if the faster R-CNN is ‘outside the comfort zone’, mostly when it works with noisy data and data that are significantly different from those used in the training stage. The validation is performed on publicly available datasets: MNIST [1] and PASCAL [2] and demonstrates that the proposed framework can be used for real-time monitoring of supervised classifiers.","PeriodicalId":226850,"journal":{"name":"2021 IEEE 24th International Conference on Information Fusion (FUSION)","volume":"127 10 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128027468","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：