search

AlfredXiangWu / face_verification_experiment

Original Caffe Version for LightCNN-9. Highly recommend to use PyTorch Version (https://github.com/AlfredXiangWu/LightCNN)
725 stars 325 forks source link

Dataset Size for Training set and Testing set of my own dataset #96

Closed psakamoori closed 7 years ago

psakamoori commented 7 years ago

Hi,

I see below table with size's picked for CASIA-WebFace training and lfw testing data set

Dataset size ec_mc_y ec_y Training set (CASIA) 144x144 48 48 Testing set (LFW) 128x128 48 40

How to choose these sizes for my own dataset (Training and Testing)? What should be size of "validation" samples during training? Any help here...

AlfredXiangWu commented 7 years ago

The size of input for the light CNN is *128128*. The 144144 for training set aims to use "crop" data augmentation. The ec_mc_y and ec_y is the parameters of face alignment we used for training and testing.

You can choose other alignment methods to normalize your face images.