Deep Learning Face Representations With Different Loss .

2y ago
10 Views
3 Downloads
5.80 MB
38 Pages
Last View : 1m ago
Last Download : 2m ago
Upload by : Aiyana Dorn
Transcription

Deep Learning Face Representationswith Different Loss Functions forFace RecognitionHongsheng LiDept. Electronic EngineeringChinese University of Hong Kong

Agenda Preview of several evaluation datasets of face recognition Preview of several different metrics Some Euclidean metric based losses Some Cosine metric based losses

Agenda Preview of several evaluation datasets of face recognition Preview of several different metrics Some Euclidean metric based losses Some Cosine metric based losses

LFW Face Verification Protocol Labeled faces in the wild (LFW) dataset is a widely used face verification(1:1) protocol, which contains 6,000 face pairs. In all 6,000 pairs, matchand mismatch pairs each account for half.Match PairsBenjamin NetanyahuMismatch PairsBarbara Felt MillerLeticia DoleraHuang, Gary B., et al. "Labeled faces in the wild: A database forstudying face recognition in unconstrained environments." Workshop onfaces in'Real-Life'Images: detection, alignment, and recognition. 2008.

MegaFace 1M Face Identification Protocol The MegaFace identification dataset includes 1M images of 690K differentindividuals (from Flickr) as the gallery set and 100K photos of 530 uniqueindividuals from FaceScrub as the probe set.Radom sample of MegaFace Photos with provided detections in redKemelmacher-Shlizerman, Ira, et al. "The megaface benchmark: 1 million faces for recognition at scale." Proceedings of the IEEEConference on Computer Vision and Pattern Recognition. 2016.Ng, Hong-Wei, and Stefan Winkler. "A data-driven approach to cleaning large face datasets." 2014 IEEE International Conference on ImageProcessing (ICIP). IEEE, 2014.

Evaluation Criterion 1:1 face verification: Accuracy 1:N face identification Accuracy with large number of distractors True Accept Rate (TAR) @ False Accept Rate (VAR) (e.g., 99.06% @ 10-2)

Agenda Preview of several evaluation datasets of face recognition Preview of several different metrics for training Some Euclidean metric based losses Some Cosine metric based losses

Preview of several different metrics fortraining Loss functions utilize metrics to evaluate the distance between outputs andtarget classes (or samples). Three different types of metrics are widely usedin losses in learning deep face representation.For a metric, the following conditions aresatisfied:Euclidean MetricCosine MetricInner Product

Agenda Preview of several evaluation datasets of face recognition Preview of several different metrics Some Euclidean metric based losses Some Cosine metric based losses

Some Euclidean metric based losses Deep learning face representation by joint identification-verification (NIPS2014) FaceNet: A unified embedding for face recognition and clustering (CVPR2015) A discriminative feature learning approach for deep face recognition (ECCV2016)

DeepID2 The key challenge of face recognition is to develop effective feature representations forreducing intra-personal variations while enlarging inter-personal differences. DeepID2uses both face identification and verification signals as loss functions.Sun, Yi, Xiaogang Wang, and Xiaoou Tang. "Deep learning face representation from predicting 10,000 classes." CVPR. 2014.Sun, Yi, et al. "Deep learning face representation by joint identification-verification." NIPS. 2014.

DeepID2Identification Loss:Softmax classifierVerification Loss:same persondifferent personSun, Yi, Xiaogang Wang, and Xiaoou Tang. "Deep learning face representation from predicting 10,000 classes." CVPR. 2014.Sun, Yi, et al. "Deep learning face representation by joint identification-verification." NIPS. 2014.

DeepID2 Some results of DeepID2Accuracy comparison with the previousbest results on LFW at that time.Sun, Yi, Xiaogang Wang, and Xiaoou Tang. "Deep learning face representation from predicting 10,000 classes." CVPR. 2014.Sun, Yi, et al. "Deep learning face representation by joint identification-verification." NIPS. 2014.

Triplet Loss (Google FaceNet) The network consists of a batchinput layer and a deep CNNfollowed by L2 normalization, whichresults in the face embedding. The Triplet Loss minimizes thedistance between an anchor and apositive, both of which have thesame identity, and maximizes thedistance between the anchor and anegative of a different identity.Schroff, Florian, Dmitry Kalenichenko, and James Philbin. "Facenet: A unified embedding for face recognition and clustering." Proceedings ofthe IEEE conference on computer vision and pattern recognition. 2015.

Triplet Loss (Google FaceNet)Schroff, Florian, Dmitry Kalenichenko, and James Philbin. "FaceNet: A unified embedding for face recognition and clustering." Proceedings ofthe IEEE conference on computer vision and pattern recognition. 2015.

Triplet Loss (Google FaceNet) MNIST feature distribution of training with triplet lossTrainingSchroff, Florian, Dmitry Kalenichenko, and James Philbin. "Facenet: A unified embedding for face recognition and clustering." Proceedings ofthe IEEE conference on computer vision and pattern recognition. 2015.

Center Loss As training approaches, DeepID2 and Triplet Loss respectively construct lossfunctions for image pairs and triplet. However, compared to the image samples, thenumber of training pairs or triplets dramatically grows. It inevitably results in slowconvergence and instability. Center loss also uses identification loss (softmax cross-entropy loss) as one oftraining supervisory signals: Besides softmax loss, an auxiliary loss item is added to gather features in theircorresponding centers:Wen, Yandong, et al. "A discriminative feature learning approach for deep face recognition." European conference on computer vision.Springer, Cham, 2016.

Center Loss Feature visualization of softmax and center lossSoftmax Loss onlySoftmax combined with CenterLossWen, Yandong, et al. "A discriminative feature learning approach for deep face recognition." European conference on computer vision.Springer, Cham, 2016.

Center Loss Some results:Verification performance of different methods onLFW and YTF datasetsIdentification rates of different methods onMegaFace with 1M distractorsWen, Yandong, et al. "A discriminative feature learning approach for deep face recognition." European conference on computer vision.Springer, Cham, 2016.

Agenda Preview of several evaluation datasets of face recognition Preview of several different metrics Some Euclidean metric based losses Some Cosine metric based losses

Some Cosine metric based losses Large-Margin Softmax Loss for Convolutional Neural Networks (ICML 2016) SphereFace: Deep Hypersphere Embedding for Face Recognition (CVPR 2017) NormFace: L2 Hypersphere Embedding for Face Verification (ACM-MM 2017) CosFace: Large Margin Cosine Loss for Deep Face Recognition (CVPR 2018) ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR2019) AdaCos: Adaptively Scaling Cosine Logit for Learning Deep FaceRepresentation (CVPR 2019)

L-Softmax L-softmax adds a marginhyperparameter in the followingformulation:Liu, Weiyang, et al. "Large-Margin Softmax Loss for Convolutional Neural Networks." ICML. Vol. 2. No. 3. 2016.

L-Softmaxψ(θ) for softmax loss and L-Softmax loss.Liu, Weiyang, et al. "Large-Margin Softmax Loss for Convolutional Neural Networks." ICML. Vol. 2. No. 3. 2016.

L-SoftmaxSome results of L-SoftmaxVerification performance (%) on LFWdataset.Liu, Weiyang, et al. "Large-Margin Softmax Loss for Convolutional Neural Networks." ICML. Vol. 2. No. 3. 2016.

A-Softmax (SphereFace) Compared with L-Softmax, A-Softmax normalized all class weights so thatmaps them into a hypersphere. This can make cosine metric more nature.Liu, Weiyang, et al. "Sphereface: Deep hypersphere embedding for face recognition." Proceedings of the IEEE conference on computervision and pattern recognition. 2017.

A-Softmax (SphereFace) Visualization of features learned with different m by using a 6class subset of the CASIA-WebFace dataset.Liu, Weiyang, et al. "Sphereface: Deep hypersphere embedding for face recognition." Proceedings of the IEEE conference on computervision and pattern recognition. 2017.

A-Softmax (SphereFace)Performance (%) on MegaFace challenge.Accuracy (%) on LFW and YTF with different number ofconvolutional layers. Left side is for LFW, while right side isfor YTF.Accuracy (%) on LFW and YTF dataset.Liu, Weiyang, et al. "Sphereface: Deep hypersphere embedding for face recognition." Proceedings of the IEEE conference on computervision and pattern recognition. 2017.

NormFace NormFace normalized both features and class weights in softmax crossentropy loss so that all logits in softmax become cosine metirc.Wang, Feng, et al. "NormFace: l 2 hypersphere embedding for face verification." Proceedings of the 25th ACM international conference onMultimedia. ACM, 2017.

NormFaceMNIST 2-D featurevisualization.The normalization operation and itsgradient in 2-dimensional space.Wang, Feng, et al. "Normface: l 2 hypersphere embedding for face verification." Proceedings of the 25th ACM international conference onMultimedia. ACM, 2017.

NormFaceLFW resutls. Here normalization indicate w/oNormFace operation.YTF resutls. Here normalization indicate w/oNormFace operation.Wang, Feng, et al. "Normface: l 2 hypersphere embedding for face verification." Proceedings of the 25th ACM international conference onMultimedia. ACM, 2017.

CosFace & ArcFace Both CosFace and ArcFace add a margin hyperparameter in NormFace.Their difference is the location where margins are added. For CosFace,margin is added on cosine metrics: For ArcFace, margin is added on angular: Although their formulations are different, their main ideas are same.Wang, Hao, et al. "Cosface: Large margin cosine loss for deep face recognition." Proceedings of the IEEE CVPR. 2018.Deng, Jiankang, et al. "Arcface: Additive angular margin loss for deep face recognition." Proceedings of the IEEE CVPR. 2019.

CosFace & ArcFaceDecision margins of CosFace andArcFace under binary classification case.Geometrical interpretation of ArcFace.A toy experiment of CosFace with different margin on 8 identities with2D features. The first row maps the 2D features onto the Euclideanspace, while the second row projects the 2D features onto the angularspace. The gap becomes evident as the margin term marginincreasesWang, Hao, et al. "Cosface: Large margin cosine loss for deep face recognition." Proceedings of the IEEE CVPR. 2018.Deng, Jiankang, et al. "Arcface: Additive angular margin loss for deep face recognition." Proceedings of the IEEE CVPR. 2019.

CosFace & ArcFaceComparison of the proposed CosFace withstate-of-the-art loss functions in face recognitioncommunity.CosFace identification and verification evaluationon MegaFace C2.Identification and verification results of different methods onMegaFace C1.Wang, Hao, et al. "Cosface: Large margin cosine loss for deep face recognition." Proceedings of the IEEE CVPR. 2018.Deng, Jiankang, et al. "Arcface: Additive angular margin loss for deep face recognition." Proceedings of the IEEE CVPR. 2019.

AdaCos The cosine-based softmax losses and its variants achieve greatsuccess in deep learning based face recognition. However, hyperparameter settings in these losses have significantinfluences on the optimization path as well as the final recognitionperformance. Manually tuning those hyperparameters heavily relies on userexperience and requires many training tricks.

AdaCos

AdaCos

AdaCosHere each sample/feature have a specifiedscaling parameter.

AdaCosIdentification and verification results of different methods on MegaFace C1.

Wang, Hao, et al. "Cosface: Large margin cosine loss for deep face recognition." Proceedings of the IEEE CVPR. 2018. Deng, Jiankang, et al. "Arcface: Additive angular margin loss for deep face recognition." Proceedings of the IEEE CVPR. 2019. Comparison of the proposed CosFace with state-of-the-art loss

Related Documents:

Deep Learning: Top 7 Ways to Get Started with MATLAB Deep Learning with MATLAB: Quick-Start Videos Start Deep Learning Faster Using Transfer Learning Transfer Learning Using AlexNet Introduction to Convolutional Neural Networks Create a Simple Deep Learning Network for Classification Deep Learning for Computer Vision with MATLAB

2.3 Deep Reinforcement Learning: Deep Q-Network 7 that the output computed is consistent with the training labels in the training set for a given image. [1] 2.3 Deep Reinforcement Learning: Deep Q-Network Deep Reinforcement Learning are implementations of Reinforcement Learning methods that use Deep Neural Networks to calculate the optimal policy.

Chapitre I. Les fonctions L des représentations d’Artin et leurs valeurs spéciales1 1. Représentations d’Artin1 2. Conjectures de Stark complexes4 3. Représentations d’Artin totalement paires et Conjecture de Gross-Stark11 4. Et les autres représentations d’Artin?14 5. Représentations d’

Garment Sizing Chart 37 Face Masks 38 PFL Face Masks 39 PFL Face Masks with No Magic Arch 39 PFL Laser Face Masks 40 PFL N-95 Particulate Respirator 40 AlphaAir Face Masks 41 AlphaGuard Face Masks 41 MVT Highly Breathable Face Masks 42 Microbreathe Face Masks 42 Coo

-The Past, Present, and Future of Deep Learning -What are Deep Neural Networks? -Diverse Applications of Deep Learning -Deep Learning Frameworks Overview of Execution Environments Parallel and Distributed DNN Training Latest Trends in HPC Technologies Challenges in Exploiting HPC Technologies for Deep Learning

Deep Learning Personal assistant Personalised learning Recommendations Réponse automatique Deep learning and Big data for cardiology. 4 2017 Deep Learning. 5 2017 Overview Machine Learning Deep Learning DeLTA. 6 2017 AI The science and engineering of making intelligent machines.

English teaching and Learning in Senior High, hoping to provide some fresh thoughts of deep learning in English of Senior High. 2. Deep learning . 2.1 The concept of deep learning . Deep learning was put forward in a paper namedon Qualitative Differences in Learning: I -

the face region and large background area are presented to verify. In recent years, a great deal of efforts have been made for face recognition with deep learning [5, 10, 18, 26, 8, 21, 20, 27]. Among the deep learning works, [5, 18, 8] learned features or deep metrics with the verification