SlideShare a Scribd company logo
1 of 18
19
1
Reviewer: Minha Kim
Domain Generalization via Shuffled
Style Assembly for Face Anti-Spoofing
19
2
CVPR’22
19
3
• Shuffled Style Assembly Network (SSAN)
– Domain generalization에 효과적인 SSAN 프레임워크 제안
– SSAN : “we split the complete representation into content and style ones with various
supervision. Then, a generalized feature space is obtained by resembling features
under a contrastive learning strategy.”
– Stylized feature space를 얻기 위해 ‘Style Transfer’를 활용
– Contrastive learning을 적용하여 liveness의 분류 성능을 높이면서 domain specific하게
학습되는 현상을 억제하는 방법을 제안
Abstract
19
4
• Domain Generalization
– 기존 연구들은 새로운 도메인에서 상대적으로 좋지 않은 detection 성능을 보여줌
– Unlabeled target data 를 학습하는 것은 비효과적일 수 있음
– 몇몇 도메인 일반화 연구가 진행되었지만, 대부분 BN layer을 적용하고 있음
– Batch Normalization (BN)은 global image statistics에 초점을 맞추기 때문에, local
image 속성을 무시할 수 있음
– Instance Normalization (IN)은 이미지 한장의 liveness-related texture 과 domain-
specific 정보를 추출 가능함
– Global + local 두 가지 정보들을 모두 획득하기 위해 BN + IN Normalization 을 적
용
Introduction
19
5
Related Work
• Normalization and Style Transfer
Adaptive Instance Normalization (AdaIN) :
기존에 있던 Style transfer의 속도와 성능을 개선시키기 위해 등장
한 Normalization 기법으로 제안됨.
content feature과 style feature을 이용해 다양한 stylized image 생
성이 가능
• Protocols for Face Anti-Spoofing
– OCIM is used to evaluate their domain generalization
– real-world에 적합한 train & test set 구성 (i.e., attack types,
such as print, replay, mask, makeup, waxworks)
19
6
Proposed Method
<Point>
1. AdaIN of StyleTransfer
2. Contrastive Learning
19
7
Proposed Method
Content and Style Information Aggregation
생성된 content feature을 서로 다른 도메인에서 구분할 수 없도록 하기 위해
Adversarial
GRL (Gradient Reversion Layer) :
도메인이 달라지더라도 충분히 일반화할 수 있도록
모델을 학습하려면, domain을 구분하는 성능은 낮
아지게 훈련해야함, 즉 역전파 동안 음의 부호를 곱
하여 gradient를 역전시켜 loss값을 최대화시킴
set of domain labels
the number of different data domains
19
8
Proposed Method
Shuffled Style Assembly - Combine both features
(Style Assembly Layer)
style feature
content feature affine parameters
19
9
Proposed Method
Contrastive Learning for Stylized Features
19
10
Proposed Method
Contrastive Learning for Stylized Features
round=domain 1,
square=domain 2
green=living,
red=spoofing
19
11
Proposed Method
19
12
• Data Evaluation Protocol & Metrics
Experiments
• Protocol 1. intra-dataset evaluation
all datasets are used as training and testing sets, simultaneously.
• Protocol 2. crossdomain evaluation.
P1: {D3, D4, D5, D10, D11, D12}, P2: {D1, D2, D6, D7, D8, D9}
Metrics : HTER ((FRR+FAR)/2), AUC
19
13
• Experiment with Leave-One-Out (LOO) setting on OCIM
Experiments
SSAN-M : mean value of the predicted depth map is the final score
SSAN-R : The value of the sigmoid function on living is the final score
OULU-NPU [3] (O), CASIA-MFSD [64] (C), Replay-Attack [6] (I), and MSU-MFSD [50] (M)
19
14
Experiments
Experiment on Limited Source Domains
19
15
Experiments
The results on the large-scale FAS benchmarks
19
16
Ablation Study
with stop-grad & without stop-grad
19
17
Ablation Study
Results of GradCAM & t-SNE
19
18
Conclusion
FAS(Face Anti-Spoofing)를 일반화할 수 있는 어셈블리 네트워크(SSAN) 제안
도메인 구별이 불가능하도록 Adversarial Learning을 채택한다.
Style feature의 경우, 도메인별 정보를 억제하면서 활력 관련 스타일 정보를 강조하기 위해
Contrastive Learning이 사용된다.
기존 데이터셋을 집계하여 FAS에 대한 대규모 벤치마크를 구축

More Related Content

Similar to [CVPR'22] Domain Generalization via Shuffled Style Assembly for Face Anti-Spoofing

생체 광학 데이터 분석 AI 경진대회 7위 수상작
생체 광학 데이터 분석 AI 경진대회 7위 수상작생체 광학 데이터 분석 AI 경진대회 7위 수상작
생체 광학 데이터 분석 AI 경진대회 7위 수상작DACON AI 데이콘
 
[ECCV2022] Generative Domain Adaptation for Face Anti-Spoofing
[ECCV2022] Generative Domain Adaptation for Face Anti-Spoofing[ECCV2022] Generative Domain Adaptation for Face Anti-Spoofing
[ECCV2022] Generative Domain Adaptation for Face Anti-SpoofingKIMMINHA3
 
PR-383: Solving ImageNet: a Unified Scheme for Training any Backbone to Top R...
PR-383: Solving ImageNet: a Unified Scheme for Training any Backbone to Top R...PR-383: Solving ImageNet: a Unified Scheme for Training any Backbone to Top R...
PR-383: Solving ImageNet: a Unified Scheme for Training any Backbone to Top R...Sunghoon Joo
 
네트워크 경량화 이모저모 @ 2020 DLD
네트워크 경량화 이모저모 @ 2020 DLD네트워크 경량화 이모저모 @ 2020 DLD
네트워크 경량화 이모저모 @ 2020 DLDKim Junghoon
 
위성관측 데이터 활용 강수량 산출 AI 경진대회 1위 수상작
위성관측 데이터 활용 강수량 산출 AI 경진대회 1위 수상작위성관측 데이터 활용 강수량 산출 AI 경진대회 1위 수상작
위성관측 데이터 활용 강수량 산출 AI 경진대회 1위 수상작DACON AI 데이콘
 
GAN based selfie-to-pokemon
GAN based selfie-to-pokemonGAN based selfie-to-pokemon
GAN based selfie-to-pokemonJuyongLee21
 
Semantic Image Synthesis with Spatially-Adaptive Normalization(GAUGAN, SPADE)
Semantic Image Synthesis with Spatially-Adaptive Normalization(GAUGAN, SPADE)Semantic Image Synthesis with Spatially-Adaptive Normalization(GAUGAN, SPADE)
Semantic Image Synthesis with Spatially-Adaptive Normalization(GAUGAN, SPADE)jungminchung
 
딥러닝 논문읽기 efficient netv2 논문리뷰
딥러닝 논문읽기 efficient netv2  논문리뷰딥러닝 논문읽기 efficient netv2  논문리뷰
딥러닝 논문읽기 efficient netv2 논문리뷰taeseon ryu
 
History of Vision AI
History of Vision AIHistory of Vision AI
History of Vision AITae Young Lee
 
Faster R-CNN
Faster R-CNNFaster R-CNN
Faster R-CNNrlawjdgns
 
Simple Review of Single Image Super Resolution Task
Simple Review of Single Image Super Resolution TaskSimple Review of Single Image Super Resolution Task
Simple Review of Single Image Super Resolution TaskMYEONGGYU LEE
 

Similar to [CVPR'22] Domain Generalization via Shuffled Style Assembly for Face Anti-Spoofing (11)

생체 광학 데이터 분석 AI 경진대회 7위 수상작
생체 광학 데이터 분석 AI 경진대회 7위 수상작생체 광학 데이터 분석 AI 경진대회 7위 수상작
생체 광학 데이터 분석 AI 경진대회 7위 수상작
 
[ECCV2022] Generative Domain Adaptation for Face Anti-Spoofing
[ECCV2022] Generative Domain Adaptation for Face Anti-Spoofing[ECCV2022] Generative Domain Adaptation for Face Anti-Spoofing
[ECCV2022] Generative Domain Adaptation for Face Anti-Spoofing
 
PR-383: Solving ImageNet: a Unified Scheme for Training any Backbone to Top R...
PR-383: Solving ImageNet: a Unified Scheme for Training any Backbone to Top R...PR-383: Solving ImageNet: a Unified Scheme for Training any Backbone to Top R...
PR-383: Solving ImageNet: a Unified Scheme for Training any Backbone to Top R...
 
네트워크 경량화 이모저모 @ 2020 DLD
네트워크 경량화 이모저모 @ 2020 DLD네트워크 경량화 이모저모 @ 2020 DLD
네트워크 경량화 이모저모 @ 2020 DLD
 
위성관측 데이터 활용 강수량 산출 AI 경진대회 1위 수상작
위성관측 데이터 활용 강수량 산출 AI 경진대회 1위 수상작위성관측 데이터 활용 강수량 산출 AI 경진대회 1위 수상작
위성관측 데이터 활용 강수량 산출 AI 경진대회 1위 수상작
 
GAN based selfie-to-pokemon
GAN based selfie-to-pokemonGAN based selfie-to-pokemon
GAN based selfie-to-pokemon
 
Semantic Image Synthesis with Spatially-Adaptive Normalization(GAUGAN, SPADE)
Semantic Image Synthesis with Spatially-Adaptive Normalization(GAUGAN, SPADE)Semantic Image Synthesis with Spatially-Adaptive Normalization(GAUGAN, SPADE)
Semantic Image Synthesis with Spatially-Adaptive Normalization(GAUGAN, SPADE)
 
딥러닝 논문읽기 efficient netv2 논문리뷰
딥러닝 논문읽기 efficient netv2  논문리뷰딥러닝 논문읽기 efficient netv2  논문리뷰
딥러닝 논문읽기 efficient netv2 논문리뷰
 
History of Vision AI
History of Vision AIHistory of Vision AI
History of Vision AI
 
Faster R-CNN
Faster R-CNNFaster R-CNN
Faster R-CNN
 
Simple Review of Single Image Super Resolution Task
Simple Review of Single Image Super Resolution TaskSimple Review of Single Image Super Resolution Task
Simple Review of Single Image Super Resolution Task
 

More from KIMMINHA3

[AAAI21] Self-Domain Adaptation for Face Anti-Spoofing
[AAAI21] Self-Domain Adaptation for Face Anti-Spoofing[AAAI21] Self-Domain Adaptation for Face Anti-Spoofing
[AAAI21] Self-Domain Adaptation for Face Anti-SpoofingKIMMINHA3
 
[TIFS'22] Learning Meta Pattern for Face Anti-Spoofing
[TIFS'22] Learning Meta Pattern for Face Anti-Spoofing[TIFS'22] Learning Meta Pattern for Face Anti-Spoofing
[TIFS'22] Learning Meta Pattern for Face Anti-SpoofingKIMMINHA3
 
[AAAI'23]Learning Polysemantic Spoof Trace
[AAAI'23]Learning Polysemantic Spoof Trace[AAAI'23]Learning Polysemantic Spoof Trace
[AAAI'23]Learning Polysemantic Spoof TraceKIMMINHA3
 
Architectures of Super-resolution (AI)
Architectures of Super-resolution (AI)Architectures of Super-resolution (AI)
Architectures of Super-resolution (AI)KIMMINHA3
 
[CVPRW2021]FReTAL: Generalizing Deepfake detection using Knowledge Distillati...
[CVPRW2021]FReTAL: Generalizing Deepfake detection using Knowledge Distillati...[CVPRW2021]FReTAL: Generalizing Deepfake detection using Knowledge Distillati...
[CVPRW2021]FReTAL: Generalizing Deepfake detection using Knowledge Distillati...KIMMINHA3
 
[NeuralIPS 2020]filter in filter pruning
[NeuralIPS 2020]filter in filter pruning[NeuralIPS 2020]filter in filter pruning
[NeuralIPS 2020]filter in filter pruningKIMMINHA3
 
Methods for interpreting and understanding deep neural networks
Methods for interpreting and understanding deep neural networksMethods for interpreting and understanding deep neural networks
Methods for interpreting and understanding deep neural networksKIMMINHA3
 
Meta learned Confidence for Few-shot Learning
Meta learned Confidence for Few-shot LearningMeta learned Confidence for Few-shot Learning
Meta learned Confidence for Few-shot LearningKIMMINHA3
 
“zero-shot” super-resolution using deep internal learning [CVPR2018]
“zero-shot” super-resolution using deep internal learning [CVPR2018]“zero-shot” super-resolution using deep internal learning [CVPR2018]
“zero-shot” super-resolution using deep internal learning [CVPR2018]KIMMINHA3
 
[CVPRW 2020]Real world Super-Resolution via Kernel Estimation and Noise Injec...
[CVPRW 2020]Real world Super-Resolution via Kernel Estimation and Noise Injec...[CVPRW 2020]Real world Super-Resolution via Kernel Estimation and Noise Injec...
[CVPRW 2020]Real world Super-Resolution via Kernel Estimation and Noise Injec...KIMMINHA3
 
Transferable GAN-generated Images Detection Framework.
Transferable GAN-generated Images  Detection Framework.Transferable GAN-generated Images  Detection Framework.
Transferable GAN-generated Images Detection Framework.KIMMINHA3
 
[Seminar arxiv]fake face detection via adaptive residuals extraction network
[Seminar arxiv]fake face detection via adaptive residuals extraction network [Seminar arxiv]fake face detection via adaptive residuals extraction network
[Seminar arxiv]fake face detection via adaptive residuals extraction network KIMMINHA3
 
Xception mhkim
Xception mhkimXception mhkim
Xception mhkimKIMMINHA3
 
short text large effect measuring the impact of user reviews on android app s...
short text large effect measuring the impact of user reviews on android app s...short text large effect measuring the impact of user reviews on android app s...
short text large effect measuring the impact of user reviews on android app s...KIMMINHA3
 

More from KIMMINHA3 (14)

[AAAI21] Self-Domain Adaptation for Face Anti-Spoofing
[AAAI21] Self-Domain Adaptation for Face Anti-Spoofing[AAAI21] Self-Domain Adaptation for Face Anti-Spoofing
[AAAI21] Self-Domain Adaptation for Face Anti-Spoofing
 
[TIFS'22] Learning Meta Pattern for Face Anti-Spoofing
[TIFS'22] Learning Meta Pattern for Face Anti-Spoofing[TIFS'22] Learning Meta Pattern for Face Anti-Spoofing
[TIFS'22] Learning Meta Pattern for Face Anti-Spoofing
 
[AAAI'23]Learning Polysemantic Spoof Trace
[AAAI'23]Learning Polysemantic Spoof Trace[AAAI'23]Learning Polysemantic Spoof Trace
[AAAI'23]Learning Polysemantic Spoof Trace
 
Architectures of Super-resolution (AI)
Architectures of Super-resolution (AI)Architectures of Super-resolution (AI)
Architectures of Super-resolution (AI)
 
[CVPRW2021]FReTAL: Generalizing Deepfake detection using Knowledge Distillati...
[CVPRW2021]FReTAL: Generalizing Deepfake detection using Knowledge Distillati...[CVPRW2021]FReTAL: Generalizing Deepfake detection using Knowledge Distillati...
[CVPRW2021]FReTAL: Generalizing Deepfake detection using Knowledge Distillati...
 
[NeuralIPS 2020]filter in filter pruning
[NeuralIPS 2020]filter in filter pruning[NeuralIPS 2020]filter in filter pruning
[NeuralIPS 2020]filter in filter pruning
 
Methods for interpreting and understanding deep neural networks
Methods for interpreting and understanding deep neural networksMethods for interpreting and understanding deep neural networks
Methods for interpreting and understanding deep neural networks
 
Meta learned Confidence for Few-shot Learning
Meta learned Confidence for Few-shot LearningMeta learned Confidence for Few-shot Learning
Meta learned Confidence for Few-shot Learning
 
“zero-shot” super-resolution using deep internal learning [CVPR2018]
“zero-shot” super-resolution using deep internal learning [CVPR2018]“zero-shot” super-resolution using deep internal learning [CVPR2018]
“zero-shot” super-resolution using deep internal learning [CVPR2018]
 
[CVPRW 2020]Real world Super-Resolution via Kernel Estimation and Noise Injec...
[CVPRW 2020]Real world Super-Resolution via Kernel Estimation and Noise Injec...[CVPRW 2020]Real world Super-Resolution via Kernel Estimation and Noise Injec...
[CVPRW 2020]Real world Super-Resolution via Kernel Estimation and Noise Injec...
 
Transferable GAN-generated Images Detection Framework.
Transferable GAN-generated Images  Detection Framework.Transferable GAN-generated Images  Detection Framework.
Transferable GAN-generated Images Detection Framework.
 
[Seminar arxiv]fake face detection via adaptive residuals extraction network
[Seminar arxiv]fake face detection via adaptive residuals extraction network [Seminar arxiv]fake face detection via adaptive residuals extraction network
[Seminar arxiv]fake face detection via adaptive residuals extraction network
 
Xception mhkim
Xception mhkimXception mhkim
Xception mhkim
 
short text large effect measuring the impact of user reviews on android app s...
short text large effect measuring the impact of user reviews on android app s...short text large effect measuring the impact of user reviews on android app s...
short text large effect measuring the impact of user reviews on android app s...
 

[CVPR'22] Domain Generalization via Shuffled Style Assembly for Face Anti-Spoofing

  • 1. 19 1 Reviewer: Minha Kim Domain Generalization via Shuffled Style Assembly for Face Anti-Spoofing
  • 3. 19 3 • Shuffled Style Assembly Network (SSAN) – Domain generalization에 효과적인 SSAN 프레임워크 제안 – SSAN : “we split the complete representation into content and style ones with various supervision. Then, a generalized feature space is obtained by resembling features under a contrastive learning strategy.” – Stylized feature space를 얻기 위해 ‘Style Transfer’를 활용 – Contrastive learning을 적용하여 liveness의 분류 성능을 높이면서 domain specific하게 학습되는 현상을 억제하는 방법을 제안 Abstract
  • 4. 19 4 • Domain Generalization – 기존 연구들은 새로운 도메인에서 상대적으로 좋지 않은 detection 성능을 보여줌 – Unlabeled target data 를 학습하는 것은 비효과적일 수 있음 – 몇몇 도메인 일반화 연구가 진행되었지만, 대부분 BN layer을 적용하고 있음 – Batch Normalization (BN)은 global image statistics에 초점을 맞추기 때문에, local image 속성을 무시할 수 있음 – Instance Normalization (IN)은 이미지 한장의 liveness-related texture 과 domain- specific 정보를 추출 가능함 – Global + local 두 가지 정보들을 모두 획득하기 위해 BN + IN Normalization 을 적 용 Introduction
  • 5. 19 5 Related Work • Normalization and Style Transfer Adaptive Instance Normalization (AdaIN) : 기존에 있던 Style transfer의 속도와 성능을 개선시키기 위해 등장 한 Normalization 기법으로 제안됨. content feature과 style feature을 이용해 다양한 stylized image 생 성이 가능 • Protocols for Face Anti-Spoofing – OCIM is used to evaluate their domain generalization – real-world에 적합한 train & test set 구성 (i.e., attack types, such as print, replay, mask, makeup, waxworks)
  • 6. 19 6 Proposed Method <Point> 1. AdaIN of StyleTransfer 2. Contrastive Learning
  • 7. 19 7 Proposed Method Content and Style Information Aggregation 생성된 content feature을 서로 다른 도메인에서 구분할 수 없도록 하기 위해 Adversarial GRL (Gradient Reversion Layer) : 도메인이 달라지더라도 충분히 일반화할 수 있도록 모델을 학습하려면, domain을 구분하는 성능은 낮 아지게 훈련해야함, 즉 역전파 동안 음의 부호를 곱 하여 gradient를 역전시켜 loss값을 최대화시킴 set of domain labels the number of different data domains
  • 8. 19 8 Proposed Method Shuffled Style Assembly - Combine both features (Style Assembly Layer) style feature content feature affine parameters
  • 10. 19 10 Proposed Method Contrastive Learning for Stylized Features round=domain 1, square=domain 2 green=living, red=spoofing
  • 12. 19 12 • Data Evaluation Protocol & Metrics Experiments • Protocol 1. intra-dataset evaluation all datasets are used as training and testing sets, simultaneously. • Protocol 2. crossdomain evaluation. P1: {D3, D4, D5, D10, D11, D12}, P2: {D1, D2, D6, D7, D8, D9} Metrics : HTER ((FRR+FAR)/2), AUC
  • 13. 19 13 • Experiment with Leave-One-Out (LOO) setting on OCIM Experiments SSAN-M : mean value of the predicted depth map is the final score SSAN-R : The value of the sigmoid function on living is the final score OULU-NPU [3] (O), CASIA-MFSD [64] (C), Replay-Attack [6] (I), and MSU-MFSD [50] (M)
  • 15. 19 15 Experiments The results on the large-scale FAS benchmarks
  • 16. 19 16 Ablation Study with stop-grad & without stop-grad
  • 18. 19 18 Conclusion FAS(Face Anti-Spoofing)를 일반화할 수 있는 어셈블리 네트워크(SSAN) 제안 도메인 구별이 불가능하도록 Adversarial Learning을 채택한다. Style feature의 경우, 도메인별 정보를 억제하면서 활력 관련 스타일 정보를 강조하기 위해 Contrastive Learning이 사용된다. 기존 데이터셋을 집계하여 FAS에 대한 대규모 벤치마크를 구축

Editor's Notes

  1. 원하는 Contents를 담고 있는 이미지의 feature xx 에서, 이미지의 스타일을 빼주고, 내가 입히고 싶은 Style을 더해주는 방식
  2. Shuffled Style Assembly Network(SSAN)의 전체 아키텍처.
  3. (Domain-Adversarial Neural Networks (DANN))
  4. styletransfer : https://lifeignite.tistory.com/46 adain : this method is widely used in generative tasks for texture synthesis and style transfer.
  5. From (1), the filter weights and FS are jointly optimized during training. After training, we merge I onto the filter weights W(I.e., W ← W I), and only use W during evaluating. Thus no additional cost is brought to the network when applying inference
  6. Half Total Error Rate (HTER) (Half Total Error Rate between FAR and FRR) ACER = (APCER + NPCER) / 2 False Positive Rate (FPR): FPR = FP / (FP + TN) True Positive Rate (TPR): TPR = TP / (TP + FN)
  7. 기존 FP는 여러 개의 fine-tunin을 필요로 하는 경우가 있었다.