GestaltMatcher

Access to source code and data

Please visit the download page in GestaltMatcher Database for accessing the code and data for training.

GestaltMatcher Architecture

Name	Type	Filter size / Stride	Output size
Input	Image	-	100x100x1
Conv 11	Convolution	3x3/1	100x100x32
Conv 12	Convolution	3x3/1	100x100x64
Pool1	Max pooling	2x2/2	50x50x64
Conv 21	Convolution	3x3/1	50x50x64
Conv 22	Convolution	3x3/1	50x50x128
Pool2	Max pooling	2x2/2	25x25x128
Conv 31	Convolution	3x3/1	25x25x96
Conv 32	Convolution	3x3/1	25x25x192
Pool3	Max pooling	2x2/2	13x13x192
Conv 41	Convolution	3x3/1	13x13x128
Conv 42	Convolution	3x3/1	13x13x256
Pool4	Max pooling	2x2/2	7x7x256
Conv 51	Convolution	3x3/1	7x7x160
Conv 52	Convolution	3x3/1	7x7x320
Pool5	Avg pooling	7x7/1	1x1x320
Dropout	Droupout (50%)	-	1x1x320
FC6	Fully connected	-	# Classes
Cost	Softmax	-	# Classes

* Important note #1: Every convolutional layer is followed by a batch normalization and a Relu layer
* Important note #2: This architecture is almost identical to CASIA-Webface face recognition architecture (see table from original paper . )