Test Result2

[glee1228]

parameter& Dataset :

[고정 파라미터]

Dataset : Sejong University building( in Seoul ) Image Dataset
input image size : ( 128,128,3 )
netVLAD clustering num : 7
training triplet data size : 700 ( 100 * 3 pairs images per class )
optimizer : SGD(learning rate=0.01, decay=1e-6, momentum=0.9, nesterov=True)
triplet loss margin : 1e-4

[유동 파라미터]

base model : VGG16( pretrained weights : 'imagenet')
total epochs : 15
train batch size : 64
T-SNE Result

base model : VGG16( pretrained weights : 'imagenet')
total epochs : 25
train batch size : 64
T-SNE Result

base model : xception( pretrained weights : 'imagenet')
total epochs : 15
train batch size : 64
T-SNE Result

base model : xception( pretrained weights : 'imagenet')
total epochs : 15
train batch size : 64
T-SNE Result

base model : inceptionV3( pretrained weights : 'imagenet')
total epochs : 15
train batch size : 64
T-SNE Result

base model : inceptionV3( pretrained weights : 'imagenet')
total epochs : 20
train batch size : 64
T-SNE Result

base model : densenet121( pretrained weights : 'imagenet')
total epochs : 15
train batch size : 64
T-SNE Result

base model : densenet169( pretrained weights : 'imagenet')
total epochs : 15
train batch size : 32
T-SNE Result