Pub Date : 2021-11-17DOI: 10.1109/IC-NIDC54101.2021.9660451
Xingxing Li, Yu Liu, Yumei Wang
Omnidirectional cameras are gaining popularity in VR/AR applications and autonomous driving due to their wide field of view. However, the images produced by the cameras have geometric distortions especially in the polar regions. This distortion poses a great challenge to computer vision tasks such as object detection. In this paper, we propose a CNN architecture called spherical CNN which is designed for omnidirectional images. According to the mapping relationship between the sphere and plane, our spherical CNN changes the size of convolution kernel and the locations of sampling points at different latitudes to adapt the image distortion. In order to verify the effectiveness of spherical CNN for the omnidirectional image object detection task, it is applied to detection network SSD(Single Shot MultiBox Detector). In our experiments, we achieve a 2% improvement on the mAP75 which represents the accuracy of detection. The experimental results verify that spherical CNN can improve the detection performance for omnidirectional images.
{"title":"Object Detection in Omnidirectional Images Based on Spherical CNN","authors":"Xingxing Li, Yu Liu, Yumei Wang","doi":"10.1109/IC-NIDC54101.2021.9660451","DOIUrl":"https://doi.org/10.1109/IC-NIDC54101.2021.9660451","url":null,"abstract":"Omnidirectional cameras are gaining popularity in VR/AR applications and autonomous driving due to their wide field of view. However, the images produced by the cameras have geometric distortions especially in the polar regions. This distortion poses a great challenge to computer vision tasks such as object detection. In this paper, we propose a CNN architecture called spherical CNN which is designed for omnidirectional images. According to the mapping relationship between the sphere and plane, our spherical CNN changes the size of convolution kernel and the locations of sampling points at different latitudes to adapt the image distortion. In order to verify the effectiveness of spherical CNN for the omnidirectional image object detection task, it is applied to detection network SSD(Single Shot MultiBox Detector). In our experiments, we achieve a 2% improvement on the mAP75 which represents the accuracy of detection. The experimental results verify that spherical CNN can improve the detection performance for omnidirectional images.","PeriodicalId":264468,"journal":{"name":"2021 7th IEEE International Conference on Network Intelligence and Digital Content (IC-NIDC)","volume":"37 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"134595400","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 : 2021-11-17DOI: 10.1109/IC-NIDC54101.2021.9660493
G. Qin, Yuwei Su, Weihao Qiu, Kai Lu, Huiling Zhou
The electronic probe trap OITD-PIS is a sensor designed to automatically monitor the number of pests in grain bulks by utilizing the pest biological characteristics. It uses two pairs of infrared photoelectric diodes less than 1 yuan as signal input and the specially designed circuit to detect pests entered the trap, leading to high cost performance and practical value. To improve the detection applicability and accuracy of collected signals when the pests are passing the diodes, this paper proposes one architecture for the infrared photoelectric sequence data acquisition with adaptive threshold. In the conventional data acquisition architecture, the CPU undertakes most of the work, which makes the system unable to guarantee the real-time performance, extends the response time for other tasks. Also, there is not an idle state for the CPU, therefore the power consumption of the system cannot be reduced. And this architecture is based on hardware-level implementation. The average time of communication response is about 180.8057ms and this architecture can help develop low-power devices. The designed adaptive threshold algorithm can effectively eliminate the detection error caused by the inconsistent parameters of infrared diodes; the sampling frequency and sampling length can be dynamically changed to accurately capture the voltage sequence data of stored grain pests with different shapes and sizes. The results showed that the architecture can accurately capture the voltage sequence data after photoelectric conversion when the object is passing the detection section.
{"title":"Design and Implementation of an Architecture for Infrared Photoelectric Sequence Data Acquisition with Adaptive Threshold","authors":"G. Qin, Yuwei Su, Weihao Qiu, Kai Lu, Huiling Zhou","doi":"10.1109/IC-NIDC54101.2021.9660493","DOIUrl":"https://doi.org/10.1109/IC-NIDC54101.2021.9660493","url":null,"abstract":"The electronic probe trap OITD-PIS is a sensor designed to automatically monitor the number of pests in grain bulks by utilizing the pest biological characteristics. It uses two pairs of infrared photoelectric diodes less than 1 yuan as signal input and the specially designed circuit to detect pests entered the trap, leading to high cost performance and practical value. To improve the detection applicability and accuracy of collected signals when the pests are passing the diodes, this paper proposes one architecture for the infrared photoelectric sequence data acquisition with adaptive threshold. In the conventional data acquisition architecture, the CPU undertakes most of the work, which makes the system unable to guarantee the real-time performance, extends the response time for other tasks. Also, there is not an idle state for the CPU, therefore the power consumption of the system cannot be reduced. And this architecture is based on hardware-level implementation. The average time of communication response is about 180.8057ms and this architecture can help develop low-power devices. The designed adaptive threshold algorithm can effectively eliminate the detection error caused by the inconsistent parameters of infrared diodes; the sampling frequency and sampling length can be dynamically changed to accurately capture the voltage sequence data of stored grain pests with different shapes and sizes. The results showed that the architecture can accurately capture the voltage sequence data after photoelectric conversion when the object is passing the detection section.","PeriodicalId":264468,"journal":{"name":"2021 7th IEEE International Conference on Network Intelligence and Digital Content (IC-NIDC)","volume":"27 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115569515","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}
Aiming to construct an evaluation dataset for task-oriented dialogues under slot filling task, this paper proposes a dataset construction approach based on two optimized data augmentation techniques named back-translation annotation synchronization and slot substitution. These optimized techniques perform well in reducing error annotations introduced by data augmentation and help maintain the style and difficulty of the original dataset. Besides, these techniques can be easily implemented by leveraging commercial interfaces and executing automated scripts, making the approach especially suitable for evaluation dataset construction. In experiments, MultiWOZ 2.0 was utilized as the benchmark dataset to generate new samples. The newly generated dialogues have lower error rate in annotations, and show the same evaluation capability as the original data, which verifies the feasibility of the construction approach and the effectiveness of two optimization methods.
{"title":"An Evaluation Dataset Construction Approach for Task-Oriented Dialogue","authors":"Weidong Liu, Shuo Liu, Donghui Gao, Rui Wang, Xuanfei Duan, Ling Jin","doi":"10.1109/IC-NIDC54101.2021.9660436","DOIUrl":"https://doi.org/10.1109/IC-NIDC54101.2021.9660436","url":null,"abstract":"Aiming to construct an evaluation dataset for task-oriented dialogues under slot filling task, this paper proposes a dataset construction approach based on two optimized data augmentation techniques named back-translation annotation synchronization and slot substitution. These optimized techniques perform well in reducing error annotations introduced by data augmentation and help maintain the style and difficulty of the original dataset. Besides, these techniques can be easily implemented by leveraging commercial interfaces and executing automated scripts, making the approach especially suitable for evaluation dataset construction. In experiments, MultiWOZ 2.0 was utilized as the benchmark dataset to generate new samples. The newly generated dialogues have lower error rate in annotations, and show the same evaluation capability as the original data, which verifies the feasibility of the construction approach and the effectiveness of two optimization methods.","PeriodicalId":264468,"journal":{"name":"2021 7th IEEE International Conference on Network Intelligence and Digital Content (IC-NIDC)","volume":"34 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121423474","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 : 2021-11-17DOI: 10.1109/IC-NIDC54101.2021.9660525
Gang Liu, Yi Liu, Xiaofeng Hong
Audio Event Recognize, which is about how to recognize audio events in the environment. It is receiving increased attention. With the development of technologies and hardware, deep learning has become the primary method of audio event recognition. In the convention methods, audio event recognition lacks supervised information. Thus, to learn from using the multiple information fusion to recognize audio events like the human auditory system, this paper proposes a method based on multitask learning of audio attribute classification. The attribute labels are defined by the audio production process. In the preliminary experiments, we add three kinds of audio attribute information to support network learning. Experiments show that for the ESC-50 and Urbansound8K datasets, audio attribute classification achieves higher accuracy, and recognition system performance improves obviously. This paper verified the stability of the three attributes and the effectiveness of attribute tags as auxiliary information.
{"title":"Audio Event Recognition by Multitask Learning of Audio Attribute Classification","authors":"Gang Liu, Yi Liu, Xiaofeng Hong","doi":"10.1109/IC-NIDC54101.2021.9660525","DOIUrl":"https://doi.org/10.1109/IC-NIDC54101.2021.9660525","url":null,"abstract":"Audio Event Recognize, which is about how to recognize audio events in the environment. It is receiving increased attention. With the development of technologies and hardware, deep learning has become the primary method of audio event recognition. In the convention methods, audio event recognition lacks supervised information. Thus, to learn from using the multiple information fusion to recognize audio events like the human auditory system, this paper proposes a method based on multitask learning of audio attribute classification. The attribute labels are defined by the audio production process. In the preliminary experiments, we add three kinds of audio attribute information to support network learning. Experiments show that for the ESC-50 and Urbansound8K datasets, audio attribute classification achieves higher accuracy, and recognition system performance improves obviously. This paper verified the stability of the three attributes and the effectiveness of attribute tags as auxiliary information.","PeriodicalId":264468,"journal":{"name":"2021 7th IEEE International Conference on Network Intelligence and Digital Content (IC-NIDC)","volume":"220 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116064457","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 : 2021-11-17DOI: 10.1109/IC-NIDC54101.2021.9660509
Qingchao Zeng, Jun Liu, Dongya Yang, Yichuan He, Xueyan Sun, Ruixiang Li, Fang Wang
Sport event detection is an important task in the research area of human behavior recognition. Owing to different motion models of different sport events, existing general human pose recognition methods cannot achieve high accuracy for sport events detection and counting. In this paper, we propose and implement a sport event detection and counting algorithm framework based on human skeletal information. Experimental evaluation results demonstrate that the algorithm can accurately detect the sit-up events and count the number of sit-ups with the highest average accuracy of 96%.
{"title":"Machine Learning Based Automatic Sport Event Detection and Counting","authors":"Qingchao Zeng, Jun Liu, Dongya Yang, Yichuan He, Xueyan Sun, Ruixiang Li, Fang Wang","doi":"10.1109/IC-NIDC54101.2021.9660509","DOIUrl":"https://doi.org/10.1109/IC-NIDC54101.2021.9660509","url":null,"abstract":"Sport event detection is an important task in the research area of human behavior recognition. Owing to different motion models of different sport events, existing general human pose recognition methods cannot achieve high accuracy for sport events detection and counting. In this paper, we propose and implement a sport event detection and counting algorithm framework based on human skeletal information. Experimental evaluation results demonstrate that the algorithm can accurately detect the sit-up events and count the number of sit-ups with the highest average accuracy of 96%.","PeriodicalId":264468,"journal":{"name":"2021 7th IEEE International Conference on Network Intelligence and Digital Content (IC-NIDC)","volume":"115 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122897736","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 : 2021-11-17DOI: 10.1109/IC-NIDC54101.2021.9660409
Yanting Zhang, Jin-jun Shi, Qingxiang Wang, Zijian Wang, Cairong Yan
Simultaneous localization and mapping (SLAM) is critical for robots in exploring an unknown environment. The monocular camera mounted on the robot can capture images continuously. However, the localization and mapping process may fail when there are not enough structure features observed from the moving camera on the robot. In this paper, we explore to use an external static surveillance camera to calculate the realtime pose data for the moving robot. We perform an adaptive self-localization for the robot taking advantage the joint information both from the camera on the robot and the external static surveillance camera. The localization results from this coordination are fused to solve the problem that localization may be unreliable in the SLAM. Whenever the SLAM fails, the estimated poses from the other camera can effectively help with the localization for the moving robot. We set up an environment to perform the experiments and validate the feasibility of coordinated mining of multiple cameras. The results can be beneficial for autonomous driving and the deployment of intelligent infrastructures.
{"title":"An Exploration of Moving Robot Localization Assisted with a Static Monocular Camera","authors":"Yanting Zhang, Jin-jun Shi, Qingxiang Wang, Zijian Wang, Cairong Yan","doi":"10.1109/IC-NIDC54101.2021.9660409","DOIUrl":"https://doi.org/10.1109/IC-NIDC54101.2021.9660409","url":null,"abstract":"Simultaneous localization and mapping (SLAM) is critical for robots in exploring an unknown environment. The monocular camera mounted on the robot can capture images continuously. However, the localization and mapping process may fail when there are not enough structure features observed from the moving camera on the robot. In this paper, we explore to use an external static surveillance camera to calculate the realtime pose data for the moving robot. We perform an adaptive self-localization for the robot taking advantage the joint information both from the camera on the robot and the external static surveillance camera. The localization results from this coordination are fused to solve the problem that localization may be unreliable in the SLAM. Whenever the SLAM fails, the estimated poses from the other camera can effectively help with the localization for the moving robot. We set up an environment to perform the experiments and validate the feasibility of coordinated mining of multiple cameras. The results can be beneficial for autonomous driving and the deployment of intelligent infrastructures.","PeriodicalId":264468,"journal":{"name":"2021 7th IEEE International Conference on Network Intelligence and Digital Content (IC-NIDC)","volume":"7 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123275978","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 : 2021-11-17DOI: 10.1109/IC-NIDC54101.2021.9660440
Zhuofan Huang, Yangsen Zhang, Ruixue Duan, Renjie Wang
With the rapid development of the Internet, the emergence of various malicious URLs seriously endangers the national network information security and user information security. Therefore, it is of great theoretical significance and practical value for network security to accurately identify and deal with malicious URLs. This paper proposes a research method of character level feature extraction and recognition of malicious URLs based on CNN + BiLSTM + CNN model. Based on the massive URL data sets, the parameter distribution characteristics of malicious URLs are analyzed, and the skip gram model is introduced to unsupervised train the preprocessed data sets, so as to embed the characters of URLs. Then the CNN + BiLSTM + CNN model is introduced to extract and optimize the local and temporal features of malicious URLs. The experimental results show that on the same data set, the malicious URL recognition method based on CNN + BiLSTM + CNN model has better recognition effect and higher accuracy than the traditional BiLSTM based algorithm and CNN based algorithm. The F1 value is increased to 98.14%, and the average iteration time is greatly reduced. It shows that the research method proposed in this paper has good applicability in the field of malicious URL identification for network security.
{"title":"Research on Malicious URL Identification and Analysis for Network Security","authors":"Zhuofan Huang, Yangsen Zhang, Ruixue Duan, Renjie Wang","doi":"10.1109/IC-NIDC54101.2021.9660440","DOIUrl":"https://doi.org/10.1109/IC-NIDC54101.2021.9660440","url":null,"abstract":"With the rapid development of the Internet, the emergence of various malicious URLs seriously endangers the national network information security and user information security. Therefore, it is of great theoretical significance and practical value for network security to accurately identify and deal with malicious URLs. This paper proposes a research method of character level feature extraction and recognition of malicious URLs based on CNN + BiLSTM + CNN model. Based on the massive URL data sets, the parameter distribution characteristics of malicious URLs are analyzed, and the skip gram model is introduced to unsupervised train the preprocessed data sets, so as to embed the characters of URLs. Then the CNN + BiLSTM + CNN model is introduced to extract and optimize the local and temporal features of malicious URLs. The experimental results show that on the same data set, the malicious URL recognition method based on CNN + BiLSTM + CNN model has better recognition effect and higher accuracy than the traditional BiLSTM based algorithm and CNN based algorithm. The F1 value is increased to 98.14%, and the average iteration time is greatly reduced. It shows that the research method proposed in this paper has good applicability in the field of malicious URL identification for network security.","PeriodicalId":264468,"journal":{"name":"2021 7th IEEE International Conference on Network Intelligence and Digital Content (IC-NIDC)","volume":"51 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123567672","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 : 2021-11-17DOI: 10.1109/IC-NIDC54101.2021.9660415
Ting Li, Ke Yu, Xiaofei Wu
The cyber trolls in social media have threatened users' personal rights and social order. By publishing offensive and disgusting comments on social media, cyber trolls try to shift the focus of the discussion, provoke others, and even trigger antagonistic behaviors among groups. Most of existing studies were based on English scenes. These methods mainly distinguished the cyber trolls from ordinary users according to whether the comments were offensive or not. But the studies ignored the diversity and concealment of cyber trolls, so it was difficult to identify them pertinently and finely. This paper builds a new Chinese cyber troll dataset and presents a hierarchical cyber troll detection method based on text and user behavior. Starting from the behavior motivation of cyber trolls, we divide users into two levels: inactive and active. For each level of users, this paper proposes some new behavior indicators based on the user statistical features, and selects the text features with significant influence from the comments. Next, these two types of features are input into the XGBoost model for detection. Finally, the detected cyber trolls at each level are combined as the final detection result. Experiments on our dataset show that our method is superior to other baseline methods.
{"title":"Hierarchical Cyber Troll Detection with Text and User Behavior","authors":"Ting Li, Ke Yu, Xiaofei Wu","doi":"10.1109/IC-NIDC54101.2021.9660415","DOIUrl":"https://doi.org/10.1109/IC-NIDC54101.2021.9660415","url":null,"abstract":"The cyber trolls in social media have threatened users' personal rights and social order. By publishing offensive and disgusting comments on social media, cyber trolls try to shift the focus of the discussion, provoke others, and even trigger antagonistic behaviors among groups. Most of existing studies were based on English scenes. These methods mainly distinguished the cyber trolls from ordinary users according to whether the comments were offensive or not. But the studies ignored the diversity and concealment of cyber trolls, so it was difficult to identify them pertinently and finely. This paper builds a new Chinese cyber troll dataset and presents a hierarchical cyber troll detection method based on text and user behavior. Starting from the behavior motivation of cyber trolls, we divide users into two levels: inactive and active. For each level of users, this paper proposes some new behavior indicators based on the user statistical features, and selects the text features with significant influence from the comments. Next, these two types of features are input into the XGBoost model for detection. Finally, the detected cyber trolls at each level are combined as the final detection result. Experiments on our dataset show that our method is superior to other baseline methods.","PeriodicalId":264468,"journal":{"name":"2021 7th IEEE International Conference on Network Intelligence and Digital Content (IC-NIDC)","volume":"46 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116926486","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 : 2021-09-25DOI: 10.1109/IC-NIDC54101.2021.9660560
Zheng Hu, Chuang Zhu, Gang He
Unsupervised person re-identification (Re-ID) is a promising and very challenging research problem in computer vision. Learning robust and discriminative features with unlabeled data is of central importance to Re-ID. Recently, more attention has been paid to unsupervised Re-ID algorithms based on clustered pseudo-label. However, the previous approaches did not fully exploit information of hard samples, simply using cluster centroid or all instances for contrastive learning. In this paper, we propose a Hard-sample Guided Hybrid Contrast Learning (HHCL) approach combining cluster-level loss with instance-level loss for unsupervised person Re-ID. Our approach applies cluster centroid contrastive loss to ensure that the network is updated in a more stable way. Meanwhile, introduction of a hard instance contrastive loss further mines the discriminative information. Extensive experiments on two popular large-scale Re-ID benchmarks demonstrate that our HHCL outperforms previous state-of-the-art methods and significantly improves the performance of unsupervised person Re-ID. The code of our work is available soon at https://github.com/bupt-ai-cz/HHCL-ReID
{"title":"Hard-sample Guided Hybrid Contrast Learning for Unsupervised Person Re-Identification","authors":"Zheng Hu, Chuang Zhu, Gang He","doi":"10.1109/IC-NIDC54101.2021.9660560","DOIUrl":"https://doi.org/10.1109/IC-NIDC54101.2021.9660560","url":null,"abstract":"Unsupervised person re-identification (Re-ID) is a promising and very challenging research problem in computer vision. Learning robust and discriminative features with unlabeled data is of central importance to Re-ID. Recently, more attention has been paid to unsupervised Re-ID algorithms based on clustered pseudo-label. However, the previous approaches did not fully exploit information of hard samples, simply using cluster centroid or all instances for contrastive learning. In this paper, we propose a Hard-sample Guided Hybrid Contrast Learning (HHCL) approach combining cluster-level loss with instance-level loss for unsupervised person Re-ID. Our approach applies cluster centroid contrastive loss to ensure that the network is updated in a more stable way. Meanwhile, introduction of a hard instance contrastive loss further mines the discriminative information. Extensive experiments on two popular large-scale Re-ID benchmarks demonstrate that our HHCL outperforms previous state-of-the-art methods and significantly improves the performance of unsupervised person Re-ID. The code of our work is available soon at https://github.com/bupt-ai-cz/HHCL-ReID","PeriodicalId":264468,"journal":{"name":"2021 7th IEEE International Conference on Network Intelligence and Digital Content (IC-NIDC)","volume":"6 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-09-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123039043","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 : 2020-09-13DOI: 10.1109/IC-NIDC54101.2021.9660538
Junhui Yin, Jiayan Qiu, Siqing Zhang, Zhanyu Ma, Jun Guo
Domain adaptive person re-identification (re-ID) is a challenging task due to the large discrepancy between the source domain and the target domain. To reduce the domain discrepancy, existing methods mainly attempt to generate pseudo labels for unlabeled target images by clustering algorithms. However, clustering methods tend to bring noisy labels and the rich fine-grained details in unlabeled images are not sufficiently exploited. In this paper, we seek to improve the quality of labels by capturing feature representation from multiple augmented views of unlabeled images. To this end, we propose a Self-Supervised Knowledge Distillation (SSKD) technique containing two modules, the identity learning and the soft label learning. Identity learning explores the relationship between unlabeled samples and predicts their one-hot labels by clustering to give exact information for confidently distinguished images. Soft label learning regards labels as a distribution and induces an image to be associated with several related classes for training peer network in a self-supervised manner, where the slowly evolving network is a core to obtain soft labels as a gentle constraint for reliable images. Finally, the two modules can resist label noise for re-ID by enhancing each other and systematically integrating label information from unlabeled images. Extensive experiments on several adaptation tasks demonstrate that the proposed method outperforms the current state-of-the-art approaches by large margins.
{"title":"SSKD: Self-Supervised Knowledge Distillation for Cross Domain Adaptive Person Re-Identification","authors":"Junhui Yin, Jiayan Qiu, Siqing Zhang, Zhanyu Ma, Jun Guo","doi":"10.1109/IC-NIDC54101.2021.9660538","DOIUrl":"https://doi.org/10.1109/IC-NIDC54101.2021.9660538","url":null,"abstract":"Domain adaptive person re-identification (re-ID) is a challenging task due to the large discrepancy between the source domain and the target domain. To reduce the domain discrepancy, existing methods mainly attempt to generate pseudo labels for unlabeled target images by clustering algorithms. However, clustering methods tend to bring noisy labels and the rich fine-grained details in unlabeled images are not sufficiently exploited. In this paper, we seek to improve the quality of labels by capturing feature representation from multiple augmented views of unlabeled images. To this end, we propose a Self-Supervised Knowledge Distillation (SSKD) technique containing two modules, the identity learning and the soft label learning. Identity learning explores the relationship between unlabeled samples and predicts their one-hot labels by clustering to give exact information for confidently distinguished images. Soft label learning regards labels as a distribution and induces an image to be associated with several related classes for training peer network in a self-supervised manner, where the slowly evolving network is a core to obtain soft labels as a gentle constraint for reliable images. Finally, the two modules can resist label noise for re-ID by enhancing each other and systematically integrating label information from unlabeled images. Extensive experiments on several adaptation tasks demonstrate that the proposed method outperforms the current state-of-the-art approaches by large margins.","PeriodicalId":264468,"journal":{"name":"2021 7th IEEE International Conference on Network Intelligence and Digital Content (IC-NIDC)","volume":"50 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-09-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127267844","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：