Robust Keypoint Detection and Matching on Fisheye Images by Self-Supervised Learning

Table 1

Parameters of the keypoint network. “DConv” denotes the deformable convolution. All convolutional layers are followed by batch normalization and an activation function of leaky ReLU, except the last layer in each head.

	Module (kernel size)	Channel (in, out)	Stride
Backbone	2 × DConv (3 × 3)	(3, 32)	1
	1 × MaxPool (3 × 3)	(32, 32)	2
	2 × DConv (3 × 3)	(32, 64)	1
	1 × MaxPool (3 × 3)	(64, 64)	2
	2 × DConv (3 × 3)	(64, 128)	1
	1 × MaxPool (3 × 3)	(128, 128)	2
	2 × DConv (3 × 3)	(128, 256)	1
	1 × DConv (3 × 3)	(256, 128)	1
Head 1	1 × DConv (3 × 3)	(128, 256)	1
	1 × DConv (3 × 3)	(256, 1)	1
	1 × sigmoid	(1, 1)	1
Head 2	1 × DConv (3 × 3)	(128, 256)	1
	1 × DConv (3 × 3)	(256, 2)	1
	1 × sigmoid	(2, 2)	1
Head 3	1 × DConv (3 × 3)	(128, 256)	1
	1 × DConv (3 × 3)	(256, 256)	1