Sunteți pe pagina 1din 10

Indian Institute of Science Department of Computational and Data Sciences

Bangalore, India
भारतीय विज्ञान संस्थान
बंगलौर, भारत
 

One Shot Face Recognition

Mid term presentation


Presented by,
Amitrajit Chattopadhyay,
3rd year UG, IISc
1.04.2020

©Department of Computational and Data Science, IISc, 2016


This work is licensed under a Creative Commons Attribution 4.0 International License
CDS
Department of Computational and Data
Copyright for external content used with attribution is retained by their original authors
CDS.IISc.ac.in | Department of Computational and
Data Sciences

Problem definition
▪ Three datasets , Training, support and testing.
▪ Training has disjoint set of labels from support and testing
▪ If the support set contains K labelled examples for each of C unique
classes, the target few-shot problem is called C-way K-shot.
▪ Support set is our actual training data as it shares the same label
with test set.
▪ But available data in support set is very small.
▪ So, we aim to perform meta-learning on the training set, extract
transferrable knowledge thus classify the test set more successfully.

2
CDS.IISc.ac.in | Department of Computational and
Data Sciences

Methodology

3
CDS.IISc.ac.in | Department of Computational and
Data Sciences
Network architecture:
• Images that enter are 84*84*3 RGB.
• The convolution blocks contains 64 3*3 filters. After
each convolution, a batch norm and a ReLU is
performed.
• The output size of last max pooling layer is 576. (in
diagram, H=576)
• The last two fully connected layers have size H*8
and H*1 respectively. The last layer is a sigmoid
layer to give a relations score between a reasonable
range.

4
CDS.IISc.ac.in | Department of Computational and
Data Sciences
Dataset structure and details
▪ In the small dataset that is used, as of now, there are 20 images
each for 150 people.
▪ The dataset is divided into training, validation and test sets in a
ratio of 64:16:20.
▪ The labels in training validation and test sets are mutually
exclusive.
▪ https://cswww.essex.ac.uk/mv/allfaces/index.html

5
CDS.IISc.ac.in | Department of Computational and
Data Sciences
Training
For each training iteration:
Select an episode for this training iteration
• Choose random set of 5 classes from training Data
• Choose 1 example in the training set from each of the 5 classes to form the sample set.
• Pick 15 samples from each of the 5 classes selected above which are not in the sample set to
form the query set.
• This means for example that there are 15×5+1×5 = 80 images in one training episode/minibatch
Adapt parameters
• Compute the embeddings of query and sample data
• Compute the relation scores of the data

• Adapt the parameters of the model by minimizing the objective function which is,

• Objective function is MSE loss between the episodes sample and query label match and the
relation score computed by current model.
• For this dataset, the training is done upto 10K epochs. 6
CDS.IISc.ac.in | Department of Computational and
Data Sciences

Testing and Results:


Batched 15 query images per class in each episode for evaluation one-
shot classification accuracies are computed by averaging over 600
randomly generated episodes from the test set.

The average accuracy came out to be: 99.97%

7
CDS.IISc.ac.in | Department of Computational and
Data Sciences

Future Plans
▪ Implement the model on a bigger (and much more difficult) dataset
and see how it performs.

▪ Experiment more with the size of network, number of convolutional


layers, no of filters and other hyper parameters to improve the
performance.
▪ See if it is possible to extend it to 20 way 1 shot learning model,
without much decrease in accuracy.
8
CDS.IISc.ac.in | Department of Computational and
Data Sciences

Other methods and how this is different


There are few notable papers in recent years on one-shot face recognition.

One-shot Face Recognition by Promoting Underrepresented Classes


Generative One-Shot Face Recognition
these papers did not follow the standard problem definition. Used test labels
during training.

Also, “Low-shot Face Recognition with Hybrid Classifiers”


shows promising results but haven’t explained how they train the model, and
lot of manual data manipulation is mentioned but not clearly documented.
“Each image in Novel Set is up-sampled 30 times by copy-paste”.

9
CDS.IISc.ac.in | Department of Computational and
Data Sciences

References
1. "Learning to Compare: Relation Network for Few-Shot Learning",
[2018] Flood Sung; Yongxin Yang; Li Zhang; Tao Xiang; Philip H.S.
Torr; Timothy M. Hospedales;
2. “One-shot Face Recognition by Promoting Underrepresented
Classes”, [2017] Yandong Guo, Lei Zhang;
3. “Generative One-Shot Face Recognition” [2018] Zhengming Ding,
Yandong Guo, Lei Zhang, Senior Member, IEEE, and Yun Fu, Senior
Member, IEEE
4. “Low-shot Face Recognition with Hybrid Classifiers” [2018] Yue
Wu, Hongfu Liu, Yun Fu

10

S-ar putea să vă placă și