search

hlamba28 / Offline-Signature-Verification-using-Siamese-Network

Identifying forged signatures using convolutional siamese networks implemented in Keras
47 stars 29 forks source link

Getting 50% test accuracy on Hindi Signatures #3

Open atinesh-s opened 4 years ago

atinesh-s commented 4 years ago

Hello hlamba28,

I have trained the code on 2 dataset CEDAR and BHSig260. I have created train, validation and test set with following partition

Dataset Total (# writers) Train Validation Test
CEDAR 55 45 5 5
BHSig260 (Bengali) 100 70 10 20
BHSig260 (Hindi) 160 110 15 35

There are total 276 (genuine - genuine) signature pairs and 276 (genuine - forged) signature pairs per users for all the dataset.

I am getting results as shown below

Dataset Accuracy
CEDAR 100%
BHSig260 (Bengali) 50%
BHSig260 (Hindi) 50%

Why accuracy on Bengali and Hindi signatures is low ?

You claimed Accuracy = 81.71% and Threshold = 0.55 for Hindi Signatures, but I am getting only 50% I tried changing train, validation and test size but test accuracy is stuck 50% not less not more

arpita739 commented 3 years ago

can you send me the code which is working

DungLeMTA commented 3 years ago

I used this code but it didnt run, can you send me your code?