Face Recognition: From Scratch To Hatch / Эдуард Тянтов (Mail.ru Group)
•Télécharger en tant que PPTX, PDF•
1 j'aime•2,206 vues
HighLoad++ 2017 Зал «Найроби+Касабланка», 7 ноября, 15:00 Тезисы: http://www.highload.ru/2017/abstracts/3044.html Мы разработали технологию по детекту и распознаванию лиц для продуктов компании Mail.ru, которая показывает высокие результаты на известных тестах. Технология на данный момент используется в Мобильном Облаке@Mail.ru для кластеризации фотографий по людям, а также во внутренних сервисах компании. ...
Recommandé
Recommandé
Contenu connexe
Tendances
Tendances (20)
Similaire à Face Recognition: From Scratch To Hatch / Эдуард Тянтов (Mail.ru Group)
Similaire à Face Recognition: From Scratch To Hatch / Эдуард Тянтов (Mail.ru Group) (20)
Plus de Ontico
Plus de Ontico (20)
Dernier
Dernier (20)
Face Recognition: From Scratch To Hatch / Эдуард Тянтов (Mail.ru Group)
- 1. Face Recognition: From Scratch To Hatch Tyantov Eduard, Mail.ru Group
- 2. Face Recognition in Cloud@Mail.ru Users upload photos to Cloud Backend identifies persons on photos, tags and show clusters
- 5. edges object parts (combination of edges) object models
- 8. Auxiliary task: facial landmarks – Face alignment: rotation – Goal: make it easier for Face Recognition Rotate
- 9. Train Datasets Wider – 32k images – 494k faces Celeba – 200k images, 10k persons – Landmarks, 40 binary attributes
- 10. Test Dataset: FDDB Face Detection Data Set and Benchmark – 2845 images – 5171 faces
- 11. Old school: Viola-Jones Haar Feature-based Cascade Classifiers Haar-like features eyes darker nose lighter Examples
- 12. Viola-Jones algorithm: training for each patch AdaBoost ensemble features 160k valuable 6k weighted sum Face or NotDataset
- 13. Viola-Jones algorithm: inference for each patch Stages Face Yes Yes Stage 1 Stage 2 Stage N Optimization – Features are grouped into stages – If a patch fails any stage => discard
- 14. Viola-Jones results OpenCV implementation – Fast: ~100ms on CPU FDDB results 0.45 – Not accurate
- 15. 1. Pre-trained network: extracting features New school: Region-based Convolutional Networks Faster RCNN, algorithm 1 CNN 2 RPN 4 Classifier Face ? 3 Roi-pooling Feature Maps 3 2. Region proposal network3. RoI-pooling: extract corresponding tensor4. Classifier: classes and the bounding box
- 16. Comparison: Viola-Jones vs R-FCN Results – 92% accuracy (R-FCN) FDDB results Viola-Jones (opencv) 0.45 HOG (dlib) 0.7 R-FCN 0.92 – 40ms on GPU (slow)
- 17. Face detection: how fast We need faster solution at the same accuracy! Target: < 10ms
- 18. Alternative: MTCNN 1 Different scales 2 Proposal CNN 3 Refine CNN 4 Output CNN Cascade of 3 CNN 1. Resize to different scales2. Proposal -> candidates + b-boxes 3. Refine -> calibration4. Output -> b-boxes + landmarks
- 19. Comparison: MTCNN vs R-FCN MTCNN + Faster + Landmarks - Less accurate - No batch processing Model GPU Inference FDDB Precision (100 errors) R-FCN 40 ms 92% MTCNN 17 ms 90%
- 20. TensorRT
- 21. What is TensorRT NVIDIA TensorRT is a high-performance deep learning inference optimizer Features – Improves performance for complex networks – FP16 & INT8 support – Effective at small batch-sizes
- 22. TensorRT: layer optimizations 2. Horizontal fusion 3. Concat elision 1. Vertical layer fusion
- 23. TensorRT: downsides 1. Caffe + TensorFlow supported 2. Fixed input/batch size 3. Basic layers support
- 24. Batch processing Problem Image size is fixed, but MTCNN works at different scales Solution Pyramid on a single image
- 25. Batch processing Results – Single run – Enables batch processing Model Inference ms MTCNN (Caffe, python) 17 MTCNN (Caffe, C++) 12.7 + batch 10.7
- 26. TensorRT: layers Problem No PReLU layer => default pre-trained model can’t be used Retrained with ReLU from scratch Model GPU Inference ms FDDB Precision (100 errors) MTCNN, batch 10.7 90% +Tensor RT 8.8 91.2% -20%
- 27. Face detection: inference Target: < 10 ms Result: 8.8 ms Ingredients 1. MTCNN 2. Batch processing 3. TensorRT
- 28. Face Recognition
- 29. Face recognition task – Goal – to compare faces Latent SpaceCNN Embedding close distant Unseen – How? To learn metric – To enable Zero-shot learning
- 30. Training set: MSCeleb – Top 100k celebrities – 10 Million images, 100 per person – Noisy: constructed by leveraging public search engines
- 31. Small test dataset: LFW Labeled Faces in the Wild Home – 13k images from the web – 1680 persons have >= 2 photos
- 32. Large test dataset: Megaface – Identification under up to 1 million “distractors” – 530 people to find
- 34. Metric Learning
- 35. Classification CNN Embedding Classify – Train CNN to predict classes – Pray for a good latent space close distant
- 36. Softmax – Learned features only separable but not discriminative – The resulting features are not sufficiently effective close
- 37. We need metric learning – Tightness of the cluster – Discriminative features
- 38. Triplet loss Features – Identity -> single point – Enforces a margin between persons Anchor Positive Negativepositive + α < negative minimize maximize
- 39. Choosing triplets Crucial problem How to choose triplets ? Useful triplets = hardest errors Pick all positive Too easy Hard enough Solution Hard-mining within a large mini-batch (>1000)
- 41. Choosing triplets: trap Anchor Positive positive ~ negative minimize maximize Negative
- 43. Choosing triplets: trap Selecting hardest negative may lead to the collapse early in training
- 44. Choosing triplets: semi-hard Pick all positive Too easy Semi-hard Too hard positive < negative < positive + α
- 45. Triplet loss: summary Overview – Requires large batches, margin tuning – Slow convergence Opensource Code – Openface (Torch) • suboptimal implementation – Facenet, not original (TensorFlow) LFW, % Megaface Openface (Torch) 92 - Our (Torch) 99.35 65 Google’s Facenet 99.63 70.5
- 46. Center loss Idea: pull the points to class centroids
- 47. Center loss: structure – Without classification loss – collapses CNN Embedding Classify Softmax Loss λ Center Loss Pull – Final loss = Softmax loss + λ Center loss
- 48. Center Loss: different lambdas λ = 10-7
- 49. Center Loss: different lambdas λ = 10-6
- 50. Center Loss: different lambdas λ = 10-5
- 51. Center loss: summary Overview – Intra-class compactness and inter-class separability – Good performance at several other tasks Opensource Code – Caffe (original, Megaface - 65%) LFW, % Megaface Triplet Loss 99.35 65 Center Loss (Torch, ours) 99.60 71.7
- 52. Tricks: augmentation Test time augmentation – Flip image Embedding Flipped Embedding Final Embedding Average – Average embeddings – Compute 2 embeddings
- 53. Tricks: alignment Rotation Kabsch algorithm - the optimal rotation matrix that minimizes the RMSD LFW, % Megaface Center Loss 99.6 71.7 Center Loss + Tricks 99.68 73
- 54. Shades on At one point we used shades augmentation How to – Construct several sunglass textures – Place them using landmarks
- 55. Tricks: adding Eye loss CNN Embedding Person Softmax Loss Center Loss – We can force CNN to learn specific discriminative features – For celebrities eye colors are available in the Internet Eye Loss Eye Color
- 56. Eye loss: summary *Adding simple features doesn’t help, i.e gender LFW, % Megaface Center Loss + Tricks 99.68 73 Center Loss + Eye 99.68 73.5
- 57. Angular Softmax ||X||= 1 On sphere Angle discriminates ||W||= 1 b=0
- 59. Angular Softmax: different «m» m=1 m=3
- 60. Angular softmax: summary Overview – As describes in the paper: doesn’t work at all – Works using sum of losses (m=1,N) over training • only on small datasets! LFW, % Megaface Center Loss 99.6 73 Center Loss + Eye 99.68 73.5 A-Softmax (Torch) 99.68 74.2 Opensource Code – Caffe (original) – Slight modification of the loss yields 74.2%
- 61. Metric learning: summary Softmax < Triplet < Center < A-Softmax A-Softmax – With bells and whistles better than center loss Center loss Overall – Rule of thumb: use Center loss – Metric learning may improve classification performance
- 62. Fighting errors
- 63. Errors after MSCeleb: children Problem Children all look alike Result Embeddings are almost single point in the space
- 64. Errors after MSCeleb: asian Problem Face Recognition’s intolerant to Asians Reason Dataset doesn’t contain enough photos of these categories
- 65. How to fix these errors ? It’s all about data, we need diverse dataset! Natural choice – avatars of social networks
- 66. A way to construct dataset Face Detection Pick largest Face Recognition+ Clustering Cleaning algorithm 1. Face detection2. Face recognition -> embeddings3. Hierarchical clustering algorithm4. Pick the largest cluster as a personIterate after each model improvement
- 67. MSCeleb dataset’s errors MSCeleb is constructed by leveraging search engines Joe Eszterhas Joe Eszterhas and Mel Gibson public confrontation leads to the error Mel Gibson =
- 70. MSCeleb dataset’s errors Dataset has been shrinked from 100k to 46k celebrities Random search engine Corrected
- 71. Results on new datasets Datasets – Train: • MSCeleb (46k) • VK-train (150k) A-Softmax on dataset Megaface MSCeleb 74.2 MSCeleb cleaned 76.2 – Test • MegaVK • Sets for children and asians MegaVK 58.4 60 + VK 77.5 87.5
- 72. Ensemble Final Embedding Concat CNN-2 CNN-1 A-Softmax model Megaface MegaVK Best single 77.5 87.5 Ensemble of 2 79.6 88.6
- 73. Workaround still … Children are still challenge for the model
- 74. Workaround Algorithm 1. Construct dataset with children 2. Compute average embedding 3. Every point inside the sphere – a child 4. Tighten distance threshold there Results This allows softening the overall threshold
- 75. How to handle big dataset It seems we can add more data infinitely, but no. Problems – Memory consumption (Softmax) – Computational costs – A lot of noise in gradients
- 76. Softmax Approximation Algorithm 1. Perform K-Means clustering using current FR model Dataset K-Means Children Women Men Smaller sets
- 77. Softmax Approximation Algorithm 1. Perform K-Means clustering using current FR model CNN Embedding Predict cluster Predict person MenPerson Softmax 2. Two Softmax heads: 1. Predicts cluster label 2. Class within the true cluster Cluster Softmax Men
- 78. Softmax Approximation Pros Push Push Harder negative 1. Prevents fusing of the clusters 2. Does hard-negative mining 3. Clusters can be specified • Children • Asian
- 79. Fighting errors on production
- 80. Errors: blur Problem • Detector yields blurry photos • Recognition forms «blurry clusters» Solution Laplacian – 2nd order derivative of the image
- 82. Errors: body parts Detection mistakes form clusters
- 83. Errors: diagrams & mushrooms
- 84. Fixing trash clusters There is similarity between “no faces”! CNN Embedding Embedding No «firing» features
- 85. Fixing trash clusters «Trash» has small norm Faces Trash Softmax loss Motivation Softmax encourage big embedding’s norms Results – ROC AUC 97% – Better then Laplacian for blurry
- 87. Fun: new governors Recently appointed governors are almost twins, but FR distinguishes them Dmitriy Gleb
- 88. Over years Face recognition algorithm captures similarity across years Although we didn’t focus on the problem
- 89. Over years
- 90. Summary 1. Use TensorRT to speed up inference 2. Metric learning: use Center loss by default 3. Clean your data thoroughly 4. Understanding CNN helps to fight errors
- 92. Auxiliary
- 93. Best avatar Problem How to pick an avatar for a person ? Solution Train model to predict awesomeness of photo
- 94. Predicting awesomeness: how to approach Social networks – not only photos, but likes too
- 95. Predicting awesomeness: dataset Awesomeness (A) = likes/audience A=18% A=27% A=75%
- 96. Results – Mean Aveage Precision @5: 25% – Data and metric are noisy => human evaluation Predicting awesomeness: summary High score Low score
- 97. Predicting awesomeness: incorporating into FR One more branch in Face Recognition CNN Small overhead awesomeness embedding face
- 98. Appendix
- 99. Histogram loss Idea – Compute similarities positive and negative pairs – Maximize a probability that a randomly sampled positive pair has smaller similarity than negative one Loss = the integral of the product between the negative distribution and the cumulative density function for the positive distribution (shown with a dashed line)
- 100. Histogram loss Results – Got nice distributions – Doesn’t improve the triplet results – LFW on 97.7%
Notes de l'éditeur
- «Detect all faces on the image at all scales and rotations»
- http://vis-www.cs.umass.edu/fddb/
- http://vis-www.cs.umass.edu/fddb/
- https://docs.opencv.org/master/d7/d8b/tutorial_py_face_detection.html , http://blog.dlib.net/2014/02/dlib-186-released-make-your-own-object.html
- R-FCN: https://arxiv.org/pdf/1605.06409.pdf , Faster: https://arxiv.org/pdf/1506.01497v3.pdf
- «Another approach – HOG features (dlib)» «not far from state-of-the-art (Google reports 94%)»
- Offtop: обучается первая сеть на 12x12, вторая – 24x24, 3 – 36x36 – hard negative mining для 2 и 3 для исправления ошибок выдает первая 5x1x1, а на пирамиде тензор, дальше составляется батч для 2 и 3 (батч над батчами не делали) https://kpzhang93.github.io/MTCNN_face_detection_alignment/paper/spl.pdf
- «Open source Pre-trained model (Caffe) No original training code «
- «but not Torch/PyTorch We’ve retrained the CNN because of PreLu absense» basic- actively devolping
- TODO: возможно показать что мы не подряд засовываем в сеть, а сразу в кучу, хз не перегруз ли
- табличка в одном месте
- https://www.microsoft.com/en-us/research/project/ms-celeb-1m-challenge-recognizing-one-million-celebrities-real-world/
- мб добавить
- Top results ~ 80% Best – 91.7%
- «All faces of person are project to a single point»
- Chad Smith
- TODO: про альфа лишнее ?
- https://ydwen.github.io/papers/WenECCV16.pdf, 65% из статьи
- TODO: убрать текст ?
- (Better than dlib)
- TODO: не выпилить но доделать нормально фотка чувака точки + очки
- «пол например не является discriminative – пробовали» 71.5% -> 75.5% или +2 71.5 vs 73, 75,5 из-за датасета
- СЛОВА «сказать что норма – это деление на сумму»
- To discriminate better -> tighten clusters
- http://proceedings.mlr.press/v48/liud16.pdf
- http://proceedings.mlr.press/v48/liud16.pdf
- “Рассказать про идею созданую по мотивам корпоратиа”
- VK – easy API OK – hard access to API, rate-limits Facebook – just closed to mining
- Cleaning algorithm Face Detection Face Recognition to get embeddings Hierarchical clustering algorithm with Soften the distance threshold Pick the largest cluster as a person iterate after each model improvement
- this leads to improved perf on megaface = 75%?
- цифра лучше ? чуть меньше текста или на 2 слайда разбить TODO: мы обнаружили что дети компанкто расположены
- https://arxiv.org/pdf/1609.04309.pdf - paper: “Efficient softmax approximation for GPUs”
- «какой профит помимо решения проблемы»
- https://docs.opencv.org/3.0-beta/doc/py_tutorials/py_imgproc/py_gradients/py_gradients.html It’s a face, but without enough facial information
- “If the detector makes a mistake, recognizer gone crazy. Detection errors often form a cluster.”
- «софтмакс стремится увеличить скалярное произведение»
- maximizing scalar product <w_i, x> for i class We normalize embeddings only for distance
- “Social networks contains avatars over users’ lives”
- Data we know How many likes Who likes
- http://people.cs.uchicago.edu/~larsson/fractalnet/overview.png
- block 6 нарисовать
- https://arxiv.org/pdf/1611.00822.pdf