Table A1.

Layers of the STN with their parameters.

Branches of the Model	Input Layer	Output Size	Filter Size/Stride	Depth
Input	Input Image	48 × 48 × 1	-	-
Localization Network	Convolution-2D	42 × 42 × 8	7 × 7/1	8
	MaxPooling-2D	21 × 21 × 8	2 × 2/2	-
	Relu	21 × 21 × 8	-	-
	Convlution-2D	17 × 17 × 10	5 × 5/1	10
	MaxPooling-2D	8 × 8 × 10	2 × 2/2	-
	Relu	8 × 8 × 10	-	-
	Fully-Connected	640	-	-
	Fully-Connected	32	-	-
	Relu	32	-	-
	Fully-Connected ($\theta$)	6	-	-
Input	Transformed Image	48 × 48 × 1	-	-
Simple-CNN	Convolution-2D	46 × 46 × 10	3 × 3/1	10
	Relu	46 × 46 × 10	-	-
	Convolution-2D	44 × 44 × 10	3 × 3/1	10
	MaxPooling-2D	22 × 22 × 10	2 × 2/2	-
	Relu	22 × 22 × 10	-	-
	Convolution-2D	20 × 20 × 10	3 × 3/1	10
	Relu	20 × 20 × 10	-	-
	Convolution-2D	18 × 18 × 10	3 × 3/1	10
	Batch Normalization	18 × 18 × 10	-	-
	MaxPooling-2D	9 × 9 × 10	2 × 2/2	-
	Relu	9 × 9 × 10	-	-
	Dropout (p = 0.5)	9 × 9 × 10	-	-
	Flatten	810	-	-
	Fully-Connected	50	-	-
	Relu	50	-	-
	Fully-Connected	8	-	-