Modern people are used to recording more and more videos using camera applications for keeping and sharing their life on social media and video-sharing platforms. To capture extensive multimedia materials, reducing the power consumption of recorded videos from camera applications plays an important role for user experience of mobile devices. This paper studies how to process and display power-saving videos recorded by camera applications on mobile devices in a real-time manner. Based on pixel-scaling methods, we design an appropriate feature map and adopt a visual attention model under the real-time limitation to effectively access attention distribution. Then, based on segmentation properties, a parallel design is appropriately applied to exploit available computation power. Next, we propose a frame-ratio predictor using machine-learning methods to efficiently predict frame ratios in a frame. Finally, the results of the comprehensive experiments conducted on a commercial smartphone with four real-world videos to evaluate the performance of the proposed design are very encouraging.
{"title":"Exploiting Machine-learning Prediction for Enabling Real-time Pixel-scaling Techniques in Mobile Camera Applications","authors":"S. Wei, Sheng-Da Tsai, Chun-Han Lin","doi":"10.1145/3555776.3577770","DOIUrl":"https://doi.org/10.1145/3555776.3577770","url":null,"abstract":"Modern people are used to recording more and more videos using camera applications for keeping and sharing their life on social media and video-sharing platforms. To capture extensive multimedia materials, reducing the power consumption of recorded videos from camera applications plays an important role for user experience of mobile devices. This paper studies how to process and display power-saving videos recorded by camera applications on mobile devices in a real-time manner. Based on pixel-scaling methods, we design an appropriate feature map and adopt a visual attention model under the real-time limitation to effectively access attention distribution. Then, based on segmentation properties, a parallel design is appropriately applied to exploit available computation power. Next, we propose a frame-ratio predictor using machine-learning methods to efficiently predict frame ratios in a frame. Finally, the results of the comprehensive experiments conducted on a commercial smartphone with four real-world videos to evaluate the performance of the proposed design are very encouraging.","PeriodicalId":42971,"journal":{"name":"Applied Computing Review","volume":null,"pages":null},"PeriodicalIF":1.0,"publicationDate":"2023-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"87027295","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}
The verification of computations performed by an untrusted server is a cornerstone for delegated computations, especially in multi-clients setting where inputs are provided by different parties. Assuming a common secret between clients, a garbled circuit offers the attractive property to ensure the correctness of a result computed by the untrusted server while keeping the input and the function private. Yet, this verification can be guaranteed only once. Based on the notion of multi-key homomorphic encryption (MKHE), we propose RMC-PVC a multi-client verifiable computation protocol, able to verify the correctness of computations performed by an untrusted server for inputs (encoded for a garbled circuit) provided by multiple clients. Thanks to MKHE, the garbled circuit is reusable an arbitrary number of times. In addition, each client can verify the computation by its own. Compared to a single-key FHE scheme, the MKHE usage in RMC-PVC allows to reduce the workload of the server and thus the response delay for the client. It also enforce the privacy of inputs, which are provided by different clients.
{"title":"RMC-PVC: A Multi-Client Reusable Verifiable Computation Protocol","authors":"Gael Marcadet, P. Lafourcade, Léo Robert","doi":"10.1145/3555776.3577680","DOIUrl":"https://doi.org/10.1145/3555776.3577680","url":null,"abstract":"The verification of computations performed by an untrusted server is a cornerstone for delegated computations, especially in multi-clients setting where inputs are provided by different parties. Assuming a common secret between clients, a garbled circuit offers the attractive property to ensure the correctness of a result computed by the untrusted server while keeping the input and the function private. Yet, this verification can be guaranteed only once. Based on the notion of multi-key homomorphic encryption (MKHE), we propose RMC-PVC a multi-client verifiable computation protocol, able to verify the correctness of computations performed by an untrusted server for inputs (encoded for a garbled circuit) provided by multiple clients. Thanks to MKHE, the garbled circuit is reusable an arbitrary number of times. In addition, each client can verify the computation by its own. Compared to a single-key FHE scheme, the MKHE usage in RMC-PVC allows to reduce the workload of the server and thus the response delay for the client. It also enforce the privacy of inputs, which are provided by different clients.","PeriodicalId":42971,"journal":{"name":"Applied Computing Review","volume":null,"pages":null},"PeriodicalIF":1.0,"publicationDate":"2023-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"87532830","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}
Shakthi Weerasinghe, A. Zaslavsky, S. Loke, A. Abken, A. Hassani, A. Medvedev
We contend that performance metrics-driven adaptive context caching has a profound impact on performance efficiency in distributed context management systems (CMS). This paper proposes an adaptive context caching approach based on (i) a model of economics-inspired expected returns of caching particular items, and (ii) learning from historical context caching performance, i.e., our approach adaptively (with respect to statistics on historical performance) caches "context" with the objective of minimizing the cost incurred by a CMS in responding to context queries. Our novel algorithm enables context queries and sub-queries to reuse and repurpose cached context in an efficient manner, different from traditional data caching. The paper also proposes heuristics and adaptive policies such as eviction and context cache memory scaling. The method is evaluated using a synthetically generated load of sub-queries inspired by a real-world scenario. We further investigate optimal adaptive caching configurations under different settings. This paper presents and discusses our findings that the proposed statistical selective caching method reaches short-term cost optimality fast under massively volatile queries. The proposed method outperforms related algorithms by up to 47.9% in cost efficiency.
{"title":"Adaptive Context Caching for Efficient Distributed Context Management Systems","authors":"Shakthi Weerasinghe, A. Zaslavsky, S. Loke, A. Abken, A. Hassani, A. Medvedev","doi":"10.1145/3555776.3577602","DOIUrl":"https://doi.org/10.1145/3555776.3577602","url":null,"abstract":"We contend that performance metrics-driven adaptive context caching has a profound impact on performance efficiency in distributed context management systems (CMS). This paper proposes an adaptive context caching approach based on (i) a model of economics-inspired expected returns of caching particular items, and (ii) learning from historical context caching performance, i.e., our approach adaptively (with respect to statistics on historical performance) caches \"context\" with the objective of minimizing the cost incurred by a CMS in responding to context queries. Our novel algorithm enables context queries and sub-queries to reuse and repurpose cached context in an efficient manner, different from traditional data caching. The paper also proposes heuristics and adaptive policies such as eviction and context cache memory scaling. The method is evaluated using a synthetically generated load of sub-queries inspired by a real-world scenario. We further investigate optimal adaptive caching configurations under different settings. This paper presents and discusses our findings that the proposed statistical selective caching method reaches short-term cost optimality fast under massively volatile queries. The proposed method outperforms related algorithms by up to 47.9% in cost efficiency.","PeriodicalId":42971,"journal":{"name":"Applied Computing Review","volume":null,"pages":null},"PeriodicalIF":1.0,"publicationDate":"2023-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"90098562","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}
In this paper we use a Twitter dataset collected between December 8, 2021 and February 18, 2022 towards the 2022 Russian invasion of Ukraine to design a data processing pipeline featuring a high accuracy Graph Convolutional Network (GCN) based political camp classifier, a botnet detection algorithm and a robust measure of botnet effects. Our experiments reveal that while the pro-Russian botnet contributes significantly to network polarization, the pro-Ukrainian botnet contributes with moderating effects. In order to understand the factors leading to different effects, we analyze the interactions between the botnets and the barrier-crossing vs. barrier-bound users on their own camps. We observe that, where as the pro-Russian botnet amplifies barrier-bound partisan users on their own camp majority of the time, the pro-Ukrainian botnet amplifies barrier-crossing users on their own camp alongside themselves majority of the time.
{"title":"Detecting and Measuring the Polarization Effects of Adversarial Botnets on Twitter","authors":"Yeonjung Lee, M. Ozer, S. Corman, H. Davulcu","doi":"10.1145/3555776.3577730","DOIUrl":"https://doi.org/10.1145/3555776.3577730","url":null,"abstract":"In this paper we use a Twitter dataset collected between December 8, 2021 and February 18, 2022 towards the 2022 Russian invasion of Ukraine to design a data processing pipeline featuring a high accuracy Graph Convolutional Network (GCN) based political camp classifier, a botnet detection algorithm and a robust measure of botnet effects. Our experiments reveal that while the pro-Russian botnet contributes significantly to network polarization, the pro-Ukrainian botnet contributes with moderating effects. In order to understand the factors leading to different effects, we analyze the interactions between the botnets and the barrier-crossing vs. barrier-bound users on their own camps. We observe that, where as the pro-Russian botnet amplifies barrier-bound partisan users on their own camp majority of the time, the pro-Ukrainian botnet amplifies barrier-crossing users on their own camp alongside themselves majority of the time.","PeriodicalId":42971,"journal":{"name":"Applied Computing Review","volume":null,"pages":null},"PeriodicalIF":1.0,"publicationDate":"2023-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"85730015","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}
Indoor Environment Quality (IEQ) is one of the most important goals for smart spaces. Thermal comfort is typically considered the most emphasized factor in IEQ that depends on personalized thermal preference. In this paper, we explore technical challenges to deploying a robot-driven personalized thermal control system that uses a mobile robot for learning user-state-specific preference efficiently. We conduct a few experiments that give a clue to overcome such challenges (i.e. low image recognition) when the system is deployed in a real world. We present future directions to improve robot-driven preference learning from the exploration.
{"title":"Towards Deployment of Mobile Robot driven Preference Learning for User-State-Specific Thermal Control in A Real-World Smart Space","authors":"Geon Kim, Hyunju Kim, Dongman Lee","doi":"10.1145/3555776.3577760","DOIUrl":"https://doi.org/10.1145/3555776.3577760","url":null,"abstract":"Indoor Environment Quality (IEQ) is one of the most important goals for smart spaces. Thermal comfort is typically considered the most emphasized factor in IEQ that depends on personalized thermal preference. In this paper, we explore technical challenges to deploying a robot-driven personalized thermal control system that uses a mobile robot for learning user-state-specific preference efficiently. We conduct a few experiments that give a clue to overcome such challenges (i.e. low image recognition) when the system is deployed in a real world. We present future directions to improve robot-driven preference learning from the exploration.","PeriodicalId":42971,"journal":{"name":"Applied Computing Review","volume":null,"pages":null},"PeriodicalIF":1.0,"publicationDate":"2023-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"86059371","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}
Learning to Rank is the task of learning a ranking function from a set of query-documents pairs. Generally, documents within a query are thousands but not all documents are informative for the learning phase. Different strategies were designed to select the most informative documents from the training set. However, most of them focused on reducing the size of the training set to speed up the learning phase, sacrificing effectiveness. A first attempt in this direction was achieved by Selective Gradient Boosting a learning algorithm that makes use of customisable sampling strategy to train effective ranking models. In this work, we propose a new sampling strategy called High_Low_Sampl for selecting negative examples applicable to Selective Gradient Boosting, without compromising model effectiveness. The proposed sampling strategy allows Selective Gradient Boosting to compose a new training set by selecting from the original one three document classes: the positive examples, high-ranked negative examples and low-ranked negative examples. The resulting dataset aims at minimizing the mis-ranking risk, i.e., enhancing the discriminative power of the learned model and maintaining generalisation to unseen instances. We demonstrated through an extensive experimental analysis on publicly available datasets, that the proposed selection algorithm is able to make the most of the negative examples within the training set and leads to models capable of obtaining statistically significant improvements in terms of NDCG, compared to the state of the art.
{"title":"On the Effect of Low-Ranked Documents: A New Sampling Function for Selective Gradient Boosting","authors":"C. Lucchese, Federico Marcuzzi, S. Orlando","doi":"10.1145/3555776.3577597","DOIUrl":"https://doi.org/10.1145/3555776.3577597","url":null,"abstract":"Learning to Rank is the task of learning a ranking function from a set of query-documents pairs. Generally, documents within a query are thousands but not all documents are informative for the learning phase. Different strategies were designed to select the most informative documents from the training set. However, most of them focused on reducing the size of the training set to speed up the learning phase, sacrificing effectiveness. A first attempt in this direction was achieved by Selective Gradient Boosting a learning algorithm that makes use of customisable sampling strategy to train effective ranking models. In this work, we propose a new sampling strategy called High_Low_Sampl for selecting negative examples applicable to Selective Gradient Boosting, without compromising model effectiveness. The proposed sampling strategy allows Selective Gradient Boosting to compose a new training set by selecting from the original one three document classes: the positive examples, high-ranked negative examples and low-ranked negative examples. The resulting dataset aims at minimizing the mis-ranking risk, i.e., enhancing the discriminative power of the learned model and maintaining generalisation to unseen instances. We demonstrated through an extensive experimental analysis on publicly available datasets, that the proposed selection algorithm is able to make the most of the negative examples within the training set and leads to models capable of obtaining statistically significant improvements in terms of NDCG, compared to the state of the art.","PeriodicalId":42971,"journal":{"name":"Applied Computing Review","volume":null,"pages":null},"PeriodicalIF":1.0,"publicationDate":"2023-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"81259953","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}
Geunsu Kim, Gyudo Park, Soohyeok Kang, Simon S. Woo
Most of the existing face recognition applications using deep learning models have leveraged CNN-based architectures as the feature extractor. However, recent studies have shown that in computer vision tasks, vision transformer-based models often outperform CNN-based models. Therefore, in this work, we propose a Sparse Vision Transformer (S-ViT) based on the Vision Transformer (ViT) architecture to improve the face recognition tasks. After the model is trained, S-ViT tends to have a sparse distribution of weights compared to ViT, so we named it according to these characteristics. Unlike the conventional ViT, our proposed S-ViT adopts image Relative Positional Encoding (iRPE) method for positional encoding. Also, S-ViT has been modified so that all token embeddings, not just class token, participate in the decoding process. Through extensive experiment, we showed that S-ViT achieves better performance in closed-set than the other baseline models, and showed better performance than the baseline ViT-based models. For example, when using ArcFace as the loss function in the identification protocol, S-ViT achieved up to 3.27% higher accuracy than ResNet50. We also show that the use of ArcFace loss functions yields greater performance gains in S-ViT than in baseline models. In addition, S-ViT has an advantage in cost-performance trade-off because it tends to be more robust to the pruning technique than the underlying model, ViT. Therefore, S-ViT offers the additional advantage, which can be applied more flexibly in the target devices with limited resources.
{"title":"S-ViT: Sparse Vision Transformer for Accurate Face Recognition","authors":"Geunsu Kim, Gyudo Park, Soohyeok Kang, Simon S. Woo","doi":"10.1145/3555776.3577640","DOIUrl":"https://doi.org/10.1145/3555776.3577640","url":null,"abstract":"Most of the existing face recognition applications using deep learning models have leveraged CNN-based architectures as the feature extractor. However, recent studies have shown that in computer vision tasks, vision transformer-based models often outperform CNN-based models. Therefore, in this work, we propose a Sparse Vision Transformer (S-ViT) based on the Vision Transformer (ViT) architecture to improve the face recognition tasks. After the model is trained, S-ViT tends to have a sparse distribution of weights compared to ViT, so we named it according to these characteristics. Unlike the conventional ViT, our proposed S-ViT adopts image Relative Positional Encoding (iRPE) method for positional encoding. Also, S-ViT has been modified so that all token embeddings, not just class token, participate in the decoding process. Through extensive experiment, we showed that S-ViT achieves better performance in closed-set than the other baseline models, and showed better performance than the baseline ViT-based models. For example, when using ArcFace as the loss function in the identification protocol, S-ViT achieved up to 3.27% higher accuracy than ResNet50. We also show that the use of ArcFace loss functions yields greater performance gains in S-ViT than in baseline models. In addition, S-ViT has an advantage in cost-performance trade-off because it tends to be more robust to the pruning technique than the underlying model, ViT. Therefore, S-ViT offers the additional advantage, which can be applied more flexibly in the target devices with limited resources.","PeriodicalId":42971,"journal":{"name":"Applied Computing Review","volume":null,"pages":null},"PeriodicalIF":1.0,"publicationDate":"2023-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"81609493","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}
D. Sloan, E. Dombay, W. Sabiiti, B. Mtafya, Ognjen Arandelovic, Marios Zachariou
Microscopy analysis of sputum images for bacilli screening is a common method used for both diagnosis and therapy monitoring of tuberculosis (TB). Nonetheless, it is a challenging procedure, since sputum examination is time-consuming and needs highly competent personnel to provide accurate results which are important for clinical decision-making. In addition, manual fluorescence microscopy examination of sputum samples for tuberculosis diagnosis and treatment monitoring is a subjective operation. In this work, we automate the process of examining fields of view (FOVs) of TB bacteria in order to determine the lipid content, and bacterial length and width. We propose a modified version of the UNet model to rapidly localise potential bacteria inside a FOV. We introduce a novel method that uses Fourier descriptors to exclude contours that do not belong to the class of bacteria, hence minimising the amount of false positives. Finally, we propose a new feature as a means of extracting a representation fed into a support vector multi-regressor in order to estimate the length and width of each bacterium. Using a real-world data corpus, the proposed method i) outperformed previous methods, and ii) estimated the cell length and width with a root mean square error of less than 0.01%.
{"title":"Estimating Phenotypic Characteristics of Tuberculosis Bacteria","authors":"D. Sloan, E. Dombay, W. Sabiiti, B. Mtafya, Ognjen Arandelovic, Marios Zachariou","doi":"10.1145/3555776.3578609","DOIUrl":"https://doi.org/10.1145/3555776.3578609","url":null,"abstract":"Microscopy analysis of sputum images for bacilli screening is a common method used for both diagnosis and therapy monitoring of tuberculosis (TB). Nonetheless, it is a challenging procedure, since sputum examination is time-consuming and needs highly competent personnel to provide accurate results which are important for clinical decision-making. In addition, manual fluorescence microscopy examination of sputum samples for tuberculosis diagnosis and treatment monitoring is a subjective operation. In this work, we automate the process of examining fields of view (FOVs) of TB bacteria in order to determine the lipid content, and bacterial length and width. We propose a modified version of the UNet model to rapidly localise potential bacteria inside a FOV. We introduce a novel method that uses Fourier descriptors to exclude contours that do not belong to the class of bacteria, hence minimising the amount of false positives. Finally, we propose a new feature as a means of extracting a representation fed into a support vector multi-regressor in order to estimate the length and width of each bacterium. Using a real-world data corpus, the proposed method i) outperformed previous methods, and ii) estimated the cell length and width with a root mean square error of less than 0.01%.","PeriodicalId":42971,"journal":{"name":"Applied Computing Review","volume":null,"pages":null},"PeriodicalIF":1.0,"publicationDate":"2023-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"85130465","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}
Liang-Chi Chen, Shu-Qi Yu, Chien-Chung Ho, Wei-Chen Wang, Yung-Chun Li
Conventional LSM tree designs delete data by inserting a delete mark to the specified key, and they thus it leaves several out-of-date values to the specified key on the LSM tree. As a result, the LSM tree encounters a serious data security issue due to the undeleted values when there arises the need for data sanitization. Sanitization is a time-consuming process that involves completely removing sensitive data from storage devices. Flash-based SSDs are widely used in many systems, but they lack an in-place update feature, which makes it difficult for LSM trees to maintain both privacy and performance on these devices. This work proposes an efficient sanitizable LSM-tree design for LSM-based key-value store over 3D NAND flash memories. Our proposed efficient sanitizable LSM-tree design focuses on integrating the processes of key-value pair updating and the execution of sanitization by exploiting our proposed influence-conscious programming method. The capability of the proposed design is evaluated by a series of experiments, for which we have very encouraging results.
{"title":"Efficient Sanitization Design for LSM-based Key-Value Store over 3D MLC NAND Flash","authors":"Liang-Chi Chen, Shu-Qi Yu, Chien-Chung Ho, Wei-Chen Wang, Yung-Chun Li","doi":"10.1145/3555776.3577780","DOIUrl":"https://doi.org/10.1145/3555776.3577780","url":null,"abstract":"Conventional LSM tree designs delete data by inserting a delete mark to the specified key, and they thus it leaves several out-of-date values to the specified key on the LSM tree. As a result, the LSM tree encounters a serious data security issue due to the undeleted values when there arises the need for data sanitization. Sanitization is a time-consuming process that involves completely removing sensitive data from storage devices. Flash-based SSDs are widely used in many systems, but they lack an in-place update feature, which makes it difficult for LSM trees to maintain both privacy and performance on these devices. This work proposes an efficient sanitizable LSM-tree design for LSM-based key-value store over 3D NAND flash memories. Our proposed efficient sanitizable LSM-tree design focuses on integrating the processes of key-value pair updating and the execution of sanitization by exploiting our proposed influence-conscious programming method. The capability of the proposed design is evaluated by a series of experiments, for which we have very encouraging results.","PeriodicalId":42971,"journal":{"name":"Applied Computing Review","volume":null,"pages":null},"PeriodicalIF":1.0,"publicationDate":"2023-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"90538526","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}
Gil-beom Lee, Jinbeom Kim, Taejune Kim, Simon S. Woo
Oriented object detection in aerial images is a challenging task due to the highly complex backgrounds and objects with arbitrary oriented and usually densely arranged. Existing oriented object detection methods adopt CNN-based methods, and they can be divided into three types: two-stage, one-stage, and anchor-free methods. All of them require non-maximum suppression (NMS) to eliminate the duplicated predictions. Recently, object detectors based on the transformer remove hand-designed components by directly solving set prediction problems via performing bipartite matching, and achieve state-of-the-art performances in general object detection. Motivated by this research, we propose a transformer-based oriented object detector named Rotated DETR with oriented bounding boxes (OBBs) labeling. We embed the scoring network to reduce the tokens corresponding to the background. In addition, we apply a proposal generator and iterative proposal refinement module in order to provide proposals with angle information to the transformer decoder. Rotated DETR achieves state-of-the-art performance on the single-stage and anchor-free oriented object detectors on DOTA, UCAS-AOD, and DIOR-R datasets with only 10% feature tokens. In the experiment, we show the effectiveness of the scoring network and iterative proposal refinement module.
{"title":"Rotated-DETR: an End-to-End Transformer-based Oriented Object Detector for Aerial Images","authors":"Gil-beom Lee, Jinbeom Kim, Taejune Kim, Simon S. Woo","doi":"10.1145/3555776.3577745","DOIUrl":"https://doi.org/10.1145/3555776.3577745","url":null,"abstract":"Oriented object detection in aerial images is a challenging task due to the highly complex backgrounds and objects with arbitrary oriented and usually densely arranged. Existing oriented object detection methods adopt CNN-based methods, and they can be divided into three types: two-stage, one-stage, and anchor-free methods. All of them require non-maximum suppression (NMS) to eliminate the duplicated predictions. Recently, object detectors based on the transformer remove hand-designed components by directly solving set prediction problems via performing bipartite matching, and achieve state-of-the-art performances in general object detection. Motivated by this research, we propose a transformer-based oriented object detector named Rotated DETR with oriented bounding boxes (OBBs) labeling. We embed the scoring network to reduce the tokens corresponding to the background. In addition, we apply a proposal generator and iterative proposal refinement module in order to provide proposals with angle information to the transformer decoder. Rotated DETR achieves state-of-the-art performance on the single-stage and anchor-free oriented object detectors on DOTA, UCAS-AOD, and DIOR-R datasets with only 10% feature tokens. In the experiment, we show the effectiveness of the scoring network and iterative proposal refinement module.","PeriodicalId":42971,"journal":{"name":"Applied Computing Review","volume":null,"pages":null},"PeriodicalIF":1.0,"publicationDate":"2023-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"89429724","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：