Pub Date : 2019-06-01DOI: 10.1109/ICB45273.2019.8987306
Myeongah Cho, Tae-Young Chung, Taeoh Kim, Sangyoun Lee
NIR-to-VIS face recognition is identifying faces of two different domains by extracting domain-invariant features. However, this is a challenging problem due to the two different domain characteristics, and the lack of NIR face dataset. In order to reduce domain discrepancy while using the existing face recognition models, we propose a ’Relation Module’ which can simply add-on to any face recognition models. The local features extracted from face image contain information of each component of the face. Based on two different domain characteristics, to use the relationships between local features is more domain-invariant than to use it as it is. In addition to these relationships, positional information such as distance from lips to chin or eye to eye, also provides domain-invariant information. In our Relation Module, Relation Layer implicitly captures relationships, and Coordinates Layer models the positional information. Also, our proposed Triplet loss with conditional margin reduces intra-class variation in training, and resulting in additional performance improvements.Different from the general face recognition models, our add-on module does not need to pre-train with the large scale dataset. The proposed module fine-tuned only with CASIA NIR-VIS 2.0 database. With the proposed module, we achieve 14.81% rank-1 accuracy and 15.47% verification rate of 0.1% FAR improvements compare to two baseline models.
{"title":"NIR-to-VIS Face Recognition via Embedding Relations and Coordinates of the Pairwise Features","authors":"Myeongah Cho, Tae-Young Chung, Taeoh Kim, Sangyoun Lee","doi":"10.1109/ICB45273.2019.8987306","DOIUrl":"https://doi.org/10.1109/ICB45273.2019.8987306","url":null,"abstract":"NIR-to-VIS face recognition is identifying faces of two different domains by extracting domain-invariant features. However, this is a challenging problem due to the two different domain characteristics, and the lack of NIR face dataset. In order to reduce domain discrepancy while using the existing face recognition models, we propose a ’Relation Module’ which can simply add-on to any face recognition models. The local features extracted from face image contain information of each component of the face. Based on two different domain characteristics, to use the relationships between local features is more domain-invariant than to use it as it is. In addition to these relationships, positional information such as distance from lips to chin or eye to eye, also provides domain-invariant information. In our Relation Module, Relation Layer implicitly captures relationships, and Coordinates Layer models the positional information. Also, our proposed Triplet loss with conditional margin reduces intra-class variation in training, and resulting in additional performance improvements.Different from the general face recognition models, our add-on module does not need to pre-train with the large scale dataset. The proposed module fine-tuned only with CASIA NIR-VIS 2.0 database. With the proposed module, we achieve 14.81% rank-1 accuracy and 15.47% verification rate of 0.1% FAR improvements compare to two baseline models.","PeriodicalId":430846,"journal":{"name":"2019 International Conference on Biometrics (ICB)","volume":"3 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"134541281","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 : 2019-06-01DOI: 10.1109/ICB45273.2019.8987261
Steven Hoffman, Renu Sharma, A. Ross
An iris recognition system is vulnerable to presentation attacks, or PAs, where an adversary presents artifacts such as printed eyes, plastic eyes or cosmetic contact lenses to defeat the system. Existing PA detection schemes do not have good generalization capability and often fail in cross-dataset scenarios, where training and testing are performed on vastly different datasets. In this work, we address this problem by fusing the outputs of three Convolutional Neural Network (CNN) based PA detectors, each of which examines different portions of the input image. The first CNN (I-CNN) focuses on the iris region only, the second CNN (F-CNN) uses the entire ocular region and the third CNN (S-CNN) uses a subset of patches sampled from the ocular region. Experiments conducted on two publicly available datasets (LivDetW15 and BERC-IF) and on a proprietary dataset (IrisID) confirm that the use of a bag of CNNs is effective in improving the generalizability of PA detectors.
{"title":"Iris + Ocular: Generalized Iris Presentation Attack Detection Using Multiple Convolutional Neural Networks","authors":"Steven Hoffman, Renu Sharma, A. Ross","doi":"10.1109/ICB45273.2019.8987261","DOIUrl":"https://doi.org/10.1109/ICB45273.2019.8987261","url":null,"abstract":"An iris recognition system is vulnerable to presentation attacks, or PAs, where an adversary presents artifacts such as printed eyes, plastic eyes or cosmetic contact lenses to defeat the system. Existing PA detection schemes do not have good generalization capability and often fail in cross-dataset scenarios, where training and testing are performed on vastly different datasets. In this work, we address this problem by fusing the outputs of three Convolutional Neural Network (CNN) based PA detectors, each of which examines different portions of the input image. The first CNN (I-CNN) focuses on the iris region only, the second CNN (F-CNN) uses the entire ocular region and the third CNN (S-CNN) uses a subset of patches sampled from the ocular region. Experiments conducted on two publicly available datasets (LivDetW15 and BERC-IF) and on a proprietary dataset (IrisID) confirm that the use of a bag of CNNs is effective in improving the generalizability of PA detectors.","PeriodicalId":430846,"journal":{"name":"2019 International Conference on Biometrics (ICB)","volume":"48 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133988154","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 : 2019-06-01DOI: 10.1109/ICB45273.2019.8987314
Xiaoyuan Wang, Li Lu, Qijun Zhao, K. Ubul
Fashion analysis has gained increasing attention thanks to its immense potential in fashion industry, precision marketing, and sociological analysis, etc. While a lot of fashion analysis work has been done for clothing and makeup, few of them address the problem from the perspective of large scale soft biometrics. In this paper, we focus on soft biometric attributes on human faces, particularly lip color and hair color, based on the analysis of which using a large scale data set we aim to reveal the fashion trend of lipstick color and hair color. To this end, we first perform the following steps on each image: face detection, occlusion detection, face parsing, and color feature extraction from the lip and hair regions. We then perform clustering based on the extracted color features in the given large scale data set. In the experiments, we collect from the Internet 15, 366 mouth-occluded and 14, 580 hair-occluded face images to train an effective occlusion detector such that noisy face images with occluded mouths/hairs are excluded from the subsequent fashion analysis, and another more than 20, 000 face images for analyzing the fashion trend of lipstick and hair colors. Our experimental results on the collected large scale data set prove the effectiveness of our proposed method.
{"title":"Hunting for Fashion via Large Scale Soft Biometrics Analysis","authors":"Xiaoyuan Wang, Li Lu, Qijun Zhao, K. Ubul","doi":"10.1109/ICB45273.2019.8987314","DOIUrl":"https://doi.org/10.1109/ICB45273.2019.8987314","url":null,"abstract":"Fashion analysis has gained increasing attention thanks to its immense potential in fashion industry, precision marketing, and sociological analysis, etc. While a lot of fashion analysis work has been done for clothing and makeup, few of them address the problem from the perspective of large scale soft biometrics. In this paper, we focus on soft biometric attributes on human faces, particularly lip color and hair color, based on the analysis of which using a large scale data set we aim to reveal the fashion trend of lipstick color and hair color. To this end, we first perform the following steps on each image: face detection, occlusion detection, face parsing, and color feature extraction from the lip and hair regions. We then perform clustering based on the extracted color features in the given large scale data set. In the experiments, we collect from the Internet 15, 366 mouth-occluded and 14, 580 hair-occluded face images to train an effective occlusion detector such that noisy face images with occluded mouths/hairs are excluded from the subsequent fashion analysis, and another more than 20, 000 face images for analyzing the fashion trend of lipstick and hair colors. Our experimental results on the collected large scale data set prove the effectiveness of our proposed method.","PeriodicalId":430846,"journal":{"name":"2019 International Conference on Biometrics (ICB)","volume":"140 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125758969","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 : 2019-06-01DOI: 10.1109/ICB45273.2019.8987425
Lázaro J. González Soler, M. Gomez-Barrero, Leonardo Chang, Airel Pérez Suárez, C. Busch
Presentation Attack Detection (PAD) is the task of determining whether a sample stems from a live subject (bona fide presentation) or from an artificial replica (Presentation Attack Instrument, PAI). Several PAD approaches have shown high effectiveness to successfully detect PAIs when the materials used for the fabrication of these PAIs are known a priori. However, most of these PAD methods do not take into account the characteristics of PAIs’ species in order to generalise to new, realistic and more challenging scenarios, where materials might be unknown. Based on that fact, in this work, we explore the impact of different PAI species, fabricated with different materials, on several local-based descriptors combined with the Fisher Vector feature encoding, in order to increase the robustness to unknown attacks. The experimental results over the well-established benchmarks of the LivDet 2011, LivDet 2013 and LivDet 2015 competitions reported error rates outperforming the top state-of-the-art in the presence of unknown attacks. Moreover, the evaluation revealed the differences in the detection performance due to the variability between the PAI species.
{"title":"On the Impact of Different Fabrication Materials on Fingerprint Presentation Attack Detection","authors":"Lázaro J. González Soler, M. Gomez-Barrero, Leonardo Chang, Airel Pérez Suárez, C. Busch","doi":"10.1109/ICB45273.2019.8987425","DOIUrl":"https://doi.org/10.1109/ICB45273.2019.8987425","url":null,"abstract":"Presentation Attack Detection (PAD) is the task of determining whether a sample stems from a live subject (bona fide presentation) or from an artificial replica (Presentation Attack Instrument, PAI). Several PAD approaches have shown high effectiveness to successfully detect PAIs when the materials used for the fabrication of these PAIs are known a priori. However, most of these PAD methods do not take into account the characteristics of PAIs’ species in order to generalise to new, realistic and more challenging scenarios, where materials might be unknown. Based on that fact, in this work, we explore the impact of different PAI species, fabricated with different materials, on several local-based descriptors combined with the Fisher Vector feature encoding, in order to increase the robustness to unknown attacks. The experimental results over the well-established benchmarks of the LivDet 2011, LivDet 2013 and LivDet 2015 competitions reported error rates outperforming the top state-of-the-art in the presence of unknown attacks. Moreover, the evaluation revealed the differences in the detection performance due to the variability between the PAI species.","PeriodicalId":430846,"journal":{"name":"2019 International Conference on Biometrics (ICB)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129304308","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}
Fingerprint recognition systems verify the identity of individuals and provide access to secure information in various commercial applications. However, with advancements in artificial intelligence, fingerprint-based security methods are vulnerable to attack. Such a breach has the potential to compromise confidential, private and valuable information. In this paper, we attack a state-of-the-art fingerprint recognition system based on transfer learning. Our approach uses attribution analysis to identify the fingerprint region crucial to correct classification, and then perturbs the fingerprint using error masks derived from a neural network to generate an adversarial fingerprint.Image quality assessment metrics applied to calculate the difference between the original and perturbed fingerprints include average difference, maximum difference, normalized absolute error, and peak signal to noise ratio. On the ATVS fingerprint dataset, the differences between these values in the original and corresponding perturbed fingerprint images are negligible. Further, the VeriFinger SDK is used to detect the minutiae and perform matching between the original and perturbed fingerprints. The matching score is above 250, which reinforces the fact that there is virtually no loss between the original and perturbed fingerprints.
{"title":"Directed Adversarial Attacks on Fingerprints using Attributions","authors":"S. Fernandes, Sunny Raj, Eddy Ortiz, Iustina Vintila, Sumit Kumar Jha","doi":"10.1109/ICB45273.2019.8987267","DOIUrl":"https://doi.org/10.1109/ICB45273.2019.8987267","url":null,"abstract":"Fingerprint recognition systems verify the identity of individuals and provide access to secure information in various commercial applications. However, with advancements in artificial intelligence, fingerprint-based security methods are vulnerable to attack. Such a breach has the potential to compromise confidential, private and valuable information. In this paper, we attack a state-of-the-art fingerprint recognition system based on transfer learning. Our approach uses attribution analysis to identify the fingerprint region crucial to correct classification, and then perturbs the fingerprint using error masks derived from a neural network to generate an adversarial fingerprint.Image quality assessment metrics applied to calculate the difference between the original and perturbed fingerprints include average difference, maximum difference, normalized absolute error, and peak signal to noise ratio. On the ATVS fingerprint dataset, the differences between these values in the original and corresponding perturbed fingerprint images are negligible. Further, the VeriFinger SDK is used to detect the minutiae and perform matching between the original and perturbed fingerprints. The matching score is above 250, which reinforces the fact that there is virtually no loss between the original and perturbed fingerprints.","PeriodicalId":430846,"journal":{"name":"2019 International Conference on Biometrics (ICB)","volume":"20 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"124419580","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 : 2019-06-01DOI: 10.1109/ICB45273.2019.8987334
R. Vicente-Garcia, Lukasz Wandzik, Louisa Grabner, J. Krüger
In this work we demonstrate the existence of demographic bias in the face representations of currently popular deep-learning-based face recognition models, exposing a bad research and development practice that may lead to a systematic discrimination of certain demographic groups in critical scenarios like automated border control. Furthermore, through the simulation of the template morphing attack, we reveal significant security risks that derive from demographic bias in current deep face models. This widely ignored problem poses important questions on fairness and accountability in face recognition.
{"title":"The Harms of Demographic Bias in Deep Face Recognition Research","authors":"R. Vicente-Garcia, Lukasz Wandzik, Louisa Grabner, J. Krüger","doi":"10.1109/ICB45273.2019.8987334","DOIUrl":"https://doi.org/10.1109/ICB45273.2019.8987334","url":null,"abstract":"In this work we demonstrate the existence of demographic bias in the face representations of currently popular deep-learning-based face recognition models, exposing a bad research and development practice that may lead to a systematic discrimination of certain demographic groups in critical scenarios like automated border control. Furthermore, through the simulation of the template morphing attack, we reveal significant security risks that derive from demographic bias in current deep face models. This widely ignored problem poses important questions on fairness and accountability in face recognition.","PeriodicalId":430846,"journal":{"name":"2019 International Conference on Biometrics (ICB)","volume":"36 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122120891","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 : 2019-06-01DOI: 10.1109/ICB45273.2019.8987387
Ruilin Li, Dehua Song, Yuhang Liu, Jufu Feng
Learning fingerprint representations is of critical importance in fingerprint indexing algorithms. Convolutional neural networks (CNNs) provide fingerprint features that perform remarkably well. In previous CNN based methods, global fingerprint features are acquired by training with entire fingerprints or by aggregating local descriptors. The former method does not make full use of the information of matched minutiae, thereby achieving relatively-low performance. While the latter way needs to extract all local features, which is time-consuming. In this paper, we propose an efficient strategy to learn global features making full use of the information of matched minutiae. We train a fully convolutional network (FCN) with local patches. Patch classes contain more information than the original fingerprint classes, and such information is helpful to learn discriminative features. In the indexing stage, we utilize the capability of FCN to get global features of whole fingerprints. Furthermore, the learned features are robust to translation, rotation, and occlusion. Therefore, we do not need to align fingerprints. The proposed approach outperforms the state-of-the-art on benchmark datasets. We achieve 99.83% identification accuracy at the penetration rate of 1% using only 256-bytes per fingerprint on NIST SD4.
{"title":"Learning Global Fingerprint Features by Training a Fully Convolutional Network with Local Patches","authors":"Ruilin Li, Dehua Song, Yuhang Liu, Jufu Feng","doi":"10.1109/ICB45273.2019.8987387","DOIUrl":"https://doi.org/10.1109/ICB45273.2019.8987387","url":null,"abstract":"Learning fingerprint representations is of critical importance in fingerprint indexing algorithms. Convolutional neural networks (CNNs) provide fingerprint features that perform remarkably well. In previous CNN based methods, global fingerprint features are acquired by training with entire fingerprints or by aggregating local descriptors. The former method does not make full use of the information of matched minutiae, thereby achieving relatively-low performance. While the latter way needs to extract all local features, which is time-consuming. In this paper, we propose an efficient strategy to learn global features making full use of the information of matched minutiae. We train a fully convolutional network (FCN) with local patches. Patch classes contain more information than the original fingerprint classes, and such information is helpful to learn discriminative features. In the indexing stage, we utilize the capability of FCN to get global features of whole fingerprints. Furthermore, the learned features are robust to translation, rotation, and occlusion. Therefore, we do not need to align fingerprints. The proposed approach outperforms the state-of-the-art on benchmark datasets. We achieve 99.83% identification accuracy at the penetration rate of 1% using only 256-bytes per fingerprint on NIST SD4.","PeriodicalId":430846,"journal":{"name":"2019 International Conference on Biometrics (ICB)","volume":"7 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115400105","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 : 2019-06-01DOI: 10.1109/ICB45273.2019.8987383
Abhishek Ranjan
ECG biometric has emerged as an appealing biometric primarily because it is difficult to spoof. Because ECG is a continuous measure of an electrophysiological signal, it is difficult to mimic, but at the same time, its day-to-day variations impact its permanence. In this paper, we present a study of the permanence of ECG biometric using a Convolutional Neural Network based authentication system and a multi-session ECG dataset collected from 800 users. The authentication system achieved an equal error rate of 2% on ECG-ID database, improving the state-of-the-art. Using this system, we designed a series of rigorous experiments by varying the days elapsed between when enrollment and authentication are performed. The results show that, despite controlling for posture, equal error rate increases as days pass. Simply including more data to enrollment does improve the accuracy, but more recent data are significantly more advantageous.
{"title":"Permanence of ECG Biometric: Experiments Using Convolutional Neural Networks","authors":"Abhishek Ranjan","doi":"10.1109/ICB45273.2019.8987383","DOIUrl":"https://doi.org/10.1109/ICB45273.2019.8987383","url":null,"abstract":"ECG biometric has emerged as an appealing biometric primarily because it is difficult to spoof. Because ECG is a continuous measure of an electrophysiological signal, it is difficult to mimic, but at the same time, its day-to-day variations impact its permanence. In this paper, we present a study of the permanence of ECG biometric using a Convolutional Neural Network based authentication system and a multi-session ECG dataset collected from 800 users. The authentication system achieved an equal error rate of 2% on ECG-ID database, improving the state-of-the-art. Using this system, we designed a series of rigorous experiments by varying the days elapsed between when enrollment and authentication are performed. The results show that, despite controlling for posture, equal error rate increases as days pass. Simply including more data to enrollment does improve the accuracy, but more recent data are significantly more advantageous.","PeriodicalId":430846,"journal":{"name":"2019 International Conference on Biometrics (ICB)","volume":"10 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122937776","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 : 2019-06-01DOI: 10.1109/ICB45273.2019.8987254
Guoqing Wang, Hu Han, S. Shan, Xilin Chen
Face recognition (FR) is being widely used in many applications from access control to smartphone unlock. As a result, face presentation attack detection (PAD) has drawn increasing attentions to secure the FR systems. Traditional approaches for PAD mainly assume that training and testing scenarios are similar in imaging conditions (illumination, scene, camera sensor, etc.), and thus may lack good generalization capability into new application scenarios. In this work, we propose an end-to-end learning approach to improve PAD generalization capability by utilizing prior knowledge from source domain via adversarial domain adaptation. We first build a source domain PAD model optimized with triplet loss. Subsequently, we perform adversarial domain adaptation w.r.t. the target domain to learn a shared embedding space by both the source and target domain models, in which the discriminator cannot reliably predict whether a sample is from the source or target domain. Finally, PAD in the target domain is performed with k-nearest neighbors (k-NN) classifier in the embedding space. The proposed approach shows promising generalization capability in a number of public-domain face PAD databases.
{"title":"Improving Cross-database Face Presentation Attack Detection via Adversarial Domain Adaptation","authors":"Guoqing Wang, Hu Han, S. Shan, Xilin Chen","doi":"10.1109/ICB45273.2019.8987254","DOIUrl":"https://doi.org/10.1109/ICB45273.2019.8987254","url":null,"abstract":"Face recognition (FR) is being widely used in many applications from access control to smartphone unlock. As a result, face presentation attack detection (PAD) has drawn increasing attentions to secure the FR systems. Traditional approaches for PAD mainly assume that training and testing scenarios are similar in imaging conditions (illumination, scene, camera sensor, etc.), and thus may lack good generalization capability into new application scenarios. In this work, we propose an end-to-end learning approach to improve PAD generalization capability by utilizing prior knowledge from source domain via adversarial domain adaptation. We first build a source domain PAD model optimized with triplet loss. Subsequently, we perform adversarial domain adaptation w.r.t. the target domain to learn a shared embedding space by both the source and target domain models, in which the discriminator cannot reliably predict whether a sample is from the source or target domain. Finally, PAD in the target domain is performed with k-nearest neighbors (k-NN) classifier in the embedding space. The proposed approach shows promising generalization capability in a number of public-domain face PAD databases.","PeriodicalId":430846,"journal":{"name":"2019 International Conference on Biometrics (ICB)","volume":"21 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125042807","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 : 2019-06-01DOI: 10.1109/ICB45273.2019.8987234
W. Tian, Feng Liu, Qijun Zhao
Existing methods for reconstructing 3D faces from multiple unconstrained images mainly focus on generating a canonical identity shape. This paper instead aims to optimize both the identity shape and the deformed shapes unique to individual images. To this end, we disentangle 3D face shapes into identity and residual components and leverage facial landmarks on the 2D images to regress both component shapes in shape space directly. Compared with existing methods, our method reconstructs more personal-ized and visually appealing 3D face shapes thanks to its ability to effectively explore both common and different shape characteristics among the multiple images and to cope with various shape deformation that is not limited to expression changes. Quantitative evaluation shows that our method achieves lower reconstruction errors than state-of-the-art methods.
{"title":"Regressing 3D Face Shapes from Arbitrary Image Sets with Disentanglement in Shape Space","authors":"W. Tian, Feng Liu, Qijun Zhao","doi":"10.1109/ICB45273.2019.8987234","DOIUrl":"https://doi.org/10.1109/ICB45273.2019.8987234","url":null,"abstract":"Existing methods for reconstructing 3D faces from multiple unconstrained images mainly focus on generating a canonical identity shape. This paper instead aims to optimize both the identity shape and the deformed shapes unique to individual images. To this end, we disentangle 3D face shapes into identity and residual components and leverage facial landmarks on the 2D images to regress both component shapes in shape space directly. Compared with existing methods, our method reconstructs more personal-ized and visually appealing 3D face shapes thanks to its ability to effectively explore both common and different shape characteristics among the multiple images and to cope with various shape deformation that is not limited to expression changes. Quantitative evaluation shows that our method achieves lower reconstruction errors than state-of-the-art methods.","PeriodicalId":430846,"journal":{"name":"2019 International Conference on Biometrics (ICB)","volume":"30 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123432494","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：