Overcoming the limits of rare disease matching using facial phenotypic descriptors

GestaltMatcher Architecture

Gestaltmatcher architecture

Name Type Filter size / Stride Output size
Conv 11Convolution3x3/1100x100x32
Conv 12Convolution3x3/1100x100x64
Pool1Max pooling2x2/250x50x64
Conv 21Convolution3x3/150x50x64
Conv 22Convolution3x3/150x50x128
Pool2Max pooling2x2/225x25x128
Conv 31Convolution3x3/125x25x96
Conv 32Convolution3x3/125x25x192
Pool3Max pooling2x2/213x13x192
Conv 41Convolution3x3/113x13x128
Conv 42Convolution3x3/113x13x256
Pool4Max pooling2x2/27x7x256
Conv 51Convolution3x3/17x7x160
Conv 52Convolution3x3/17x7x320
Pool5Avg pooling7x7/11x1x320
DropoutDroupout (50%)-1x1x320
FC6Fully connected-# Classes
CostSoftmax-# Classes
* Important note #1: Every convulotional 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 . )