Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman
Deep face recognition
Prepare for your exams
Study with the several resources on Docsity
Earn points to download
Earn points by helping other students or get them with a premium plan
Guidelines and tips
Prepare for your exams
Study with the several resources on Docsity
Earn points to download
Earn points by helping other students or get them with a premium plan
Slides from lecture on ethics and computer science
Typology: Slides
2020/2021
1 / 30
This page cannot be seen from the preview
Don't miss anything!
Related documents
Partial preview of the text
Download Deep face recognition and more Slides Computer science in PDF only on Docsity!
Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman
Deep face recognition
Key Questions
- (^) Can large scale datasets be built with minimal human intervention?
- (^) Can we propose a convolutional neural network which can compete with that of internet giants like Google and Facebook? Yes! Yes!
Can large scale datasets be built with minimal human intervention?
Dataset Collection
1. Candidate list generation : Finding names of celebrities
- (^) Tap the knowledge on the web
- (^) 5000 identities Robert Downey Jr. Allison Hannigan Ashley Hamilton Amitabh Bacchan Barack Obama Vladimir Putin
Dataset Collection
Popular Celebrity
Robert Downey Jr.
Dataset Collection
Not so popular
celebrity
Ashley Hamilton
3. Rank image sets
- (^) 2000 images per identity
- (^) Searching by appending keyword “actor”
- (^) Learning classifier using data obtained the previous step. c
- (^) Ranking 2000 images and selecting top 1000 images
- (^) Approx. 2.6 Million images of 2622 celebrities Dataset Collection
Dataset Collection
4. Near duplicate removal
- (^) VLAD descriptor based near duplicate removal
5. manual filtering
- (^) Curating the dataset further using manual
checks
**No. Aim Mode # Persons
images
/person Total # images Anno. effort** 1 Candidate list generation Auto 5000 200 1,000,000 - 2 Candidate list filtering Manual 2622 - - 4 days 3 Rank image sets Auto 2622 1000 2,622,000 - 4 Near duplicate removal Auto 2622 623 1,635,159 - 5 Manual filtering Manual 2622 375 982,803 10 days
Dataset Collection
Example Images From Our Dataset
Can we propose a convolutional neural network which can compete with that of internet giants like Google and Facebook etc.?
Convolutional Neural Network
Training
- (^) MatConvNet Tootlbox
- (^) Nvidia CuDNN bindings
- (^) Multi GPU Training (approx 3.5x speedup)
- (^) Nvidia Titan Black
- (^) 7 days of training
- (^) Random Gaussian Initialization
- (^) Stochastic Gradient Descent with back prop.
- (^) Accumulator Descent for large batch sizes
- (^) Batch Size: 256
- (^) Incremental FC layer training
- (^) 2622 way multi class criterion (soft max) image Conv- maxpool fc- fc- Softmax Conv- Conv- maxpool Conv- Conv- maxpool Conv- Conv- maxpool Conv- Conv- Conv- maxpool Conv- Conv- Matconvnet – convolutional neural networks for matlab.^ fc- A Vedaldi and K. Lenc. Arxiv - 2014.
Convolutional Neural Network
Training: Learning Task Specific Embedding
- (^) Learning embedding by minimizing triplet loss
- (^) Learning a projection from 4096 to 1024 dimensions
- (^) On line triplet formation at the beginning of each iteration
- (^) Fine tuned on target datasets
- (^) Only the projection layers learnt
Labeled Faces In the Wild Dataset (LFW)
- (^) Face Verification: Given a pair of images specify
whether they belong to the same person
- (^) 13K images, 5.7K people
- (^) Standard benchmark in the community
- (^) Several test protocols depending upon availability of training data
within and outside th
dataset. [ G. Huang, M. Ramesh, T. Berg and E. Learned-Miller - Tech Report 07]
Effects of design choices (LFW Unrestricted Protocol) No. Network Config. Dataset Face Align Training Face Align Testing Embedding 100%-EER