Sei sulla pagina 1di 21

Face recognition

Face Recognition
• Identification
• Xác định probe trong gallery
• Closed set
• Open set
• Verification
• Giả sử không biết trước người trong hai ảnh cần verify
LFW – Labeled Faces In The Wild
• Dễ sử dụng
• Unconstrained
• Protocol tin cậy để so sánh
• Có thể replicate kết quả
LFW – Labeled Faces In The Wild
• 13233 images
• 5749 people
• 1680 people with two or more images
• Sử dụng Viola – Jones detector
• 6000 cặp ảnh
LFW
• Pose variation: within 20 degrees of frontal
• Khoảng cách tuổi: 4.94±4.24 (pos pairs), 14.85±11.00 (neg pairs)
• 6000 cặp ảnh là tương đối dễ
• Bias
• Ảnh lấy trên các trang tin tức
• Nhiều người nổi tiếng
• Thực tế:
• Ảnh mờ, camera motion
• Ảnh thiếu sáng
Performance on LFW
Outside Verif.
Method Net. Loss # models Aligned Layers Accu.
data metric

DeepFace ident. 4M 4 3D wt. chi-sq. 8 97.35±0.25

Canon. view CNN ident. 203K 60 2D Jt. Bayes 7 96.45±0.25

DeepID ident. 203K 60 2D Jt. Bayes 7 97.45±0.26

DeepID2 ident. + verif. 203K 25 2D Jt. Bayes 7 99.15±0.13

DeepID2+ ident. + verif. 290K 25 2D Jt. Bayes 7 99.47±0.12

DeepID3 ident. + verif. 290K 25 2D Jt. Bayes 10-15 99.53±0.10

Face++ ident. 5M 1 2D L2 10 99.50±0.36

FaceNet verif. (triplet) 260M 1 no L2 22 99.60±0.09

Tencent - 1M 20 yes Jt. Bayes 12 99.65±0.25

Source: Labeled Faces in the Wild: A Survey


Baidu

Source: Targeting Ultimate Accuracy: Face Recognition via Deep Embedding


Baidu
Verification vs identification
• Identification: difficulity ~ number of people in the gallery
• True accept rate (TAR) and False accept rate (FAR) for open set
• Verification system 99.0% accuracy
• 100 match pairs, and 100 mismatched pairs: 2 errors
• Use that system in a closed set identification scenario with 901
subjects
• 1% of 900 mismatched to be rated as a “match”
• 1 correct identity and 9 incorrect identities > match threshold
• 10 identities are very similar
New databases and benchmarks
• IJB-A: open set identification, verification, face detection
• FaceScrub & CASIA: deep data set
• MegaFace: large scale face recognition
• Cross-Age LFW: age gaps
• MegaFace2: train on same data, test at the million scale
• YouTubeFace
IJB-A:

Source: Pushing the Frontiers of Unconstrained Face Detection and Recognition: IARPA Janus Benchmark A, 2015
FaceScrub & CASIA
• FaceScrub: 107,818 images of celebrities
• 530 individuals, 200 images/person
• CASIA: 494,414 images
• 10,575 individuals, 46 images/person
• Outliers are automatically rejected
• Limit to the degree of variability
• Difficult to know % of correct labels in the database
• Large and deep databases
MegaFace: large scale face recognition
• Large number of distractors. LFW = 10 distractors
• One image each of 1M different individuals
• Conjunction with other database (FaceScrub)
• Gallery: 1 image of a person + 1M distractors
• Probe: M – 1 images of the person
• Goal: Identify the single matching individual from among 1,000,001
identities in the gallery.
• 95% performance on LFW: 35-75% identification rates with 1M
distractors

Source: The MegaFace Benchmark: 1 Million Faces for Recognition at Scale


Cross-Age: LFW
• Positive pairs: larger age gaps
• Negative pairs: using individuals
with the same race and gender
• Accuracy drop 10% - 17%

Source: Cross-Age LFW: A Database for Studying Cross-Age


Face Recognition in
Unconstrained Environments
MegaFace2 benchmark
All algorithms to be
• trained on same data
• tested at million scale
Top 1 result (30/8/2017)
• FGNet: 61.179 (Iden.),
63.496 (verf)
• FaceScrub: 77.068 (Iden),
90.302 (verf)
• TencentAILab FaceCNN
v1

Source: Level Playing Field for Million Scale Face Recognition


YouTubeFace
• Unconstrained face recognition
in videos
• 3,425 videos of 1,595 different
people
• Matching 5000 pairs, 10 folds

Source: Face Recognition in Unconstrained Videos with Matched Background Similarity


Attention-Set based Metric Learning
• ASML: correct sample biases
• Measure the correlations among
face image sets
• End-to-end trainable CNN
assembled with ASML

Source: Attention-Set based Metric Learning for Video Face Recognition


Deep 3D Face Identification
• First to use a deep CNN for 3D face recognition
• Frontalize a 3D scan
• Generate 2.5D depth map
• Extract deep features to represent the 3D surface
• 3D face augmentation
• Enlarge a limited 3D dataset

Source: Deep 3D Face Identification


Deep Visage
• Can we achieve state-of-the-art results with a single CNN model which
is trained only once with the identity labels?
• Single CNN
• Softmax Loss
• MSCeleb
• Feature normalization (Batch Normalization)

Source: DeepVisage: Making face recognition simple yet with powerful generalization skills
Alignment
• Pose variations
• aligning the input images
• building mappings
• conditioning on pose (separate classifiers for each category of pose)
• model can learn to compare images
across pose
• FaceNet improves from 98.87% to 99.63% using a trained alignment
system.

Potrebbero piacerti anche