Siamese-Triplet Networks using Pytorch

Face Recognition is genarlly a one-shot learning task. One shot learning is a classification task where the model should learn from one example of given class and be able to recognize it in the future.

Siamese Network here is used to implement the one-shot learning for face recognition.

Dataset : The Database of Faces (AT&T)

The AT&T face dataset, “(formerly ‘The ORL Database of Faces’) is used for training face verification and recognititon model.

Dataset Statistics

1. Color: Grey-scale

2. Sample Size: 92×112

3. #Samples: 400

   There are 10 different images of each of 40 distinct subjects.

Architechtures

Parameter
Value

Training Set
75% (300/400)

Testing Set
25% (100/400)

Number of Epochs
16

Learning Rate
10-4

Total Parameters
4,170,400

Loss Function
Triplet Loss

Optimizer
Adam

Train Accuracy
92.67 %

Test Accuracy
88.0 %

Total Accuracy
87.25 %

Parameter
Face Identification
One Shot Learning

Training Set
70% (38×7/38×10)
75% (300/400)

Testing Set
30% (38×3/38×10)
25% (100/400)

Number of Epochs
8
20

Learning Rate
20-4
10-4

Total Parameters
11,235,904
11,235,904

Loss Function
Triplet Loss
Triplet Loss

Optimizer
Adam
Adam

Threshold
8
–

Train Accuracy
99.62 %
82.00 %

Test Accuracy
94.73 %
87.00 %

Total Accuracy
92.75 %
75.50 %

Parameter
Value

Training Set
75% (300/400)

Testing Set
25% (100/400)

Number of Epochs
20

Learning Rate
20-4

Total Parameters
17,728,064

Loss Function
Triplet Loss

Optimizer
Adam

Train Accuracy
93.00 %

Test Accuracy
69.00 %

Total Accuracy
82.00 %

References: