An Improved 3D Shape Recognition Method Based on Panoramic View

Table 1

Comparison of the parameters of architectures used in the three methods. In this table, “conv” represents convolutional layer and “fc” represents fully connected layer. For each convolutional layer, the first number in brackets is the kernel size, and the second is the number of kernels. For each fully connected layer, the number represents the number of hidden units in this layer.

Method	3D ShapeNets [7]	DeepPano [21]	Our Method
Input Size	30 × 30 × 30	160 × 64	108 × 36
Conv1	(6, 48)	(5, 96)	(1, 64)
Conv2	(5, 160)	(5, 256)	(2, 80)
Conv3	(4, 512)	(3, 384)	(4, 160)
Conv4	None	(3, 512)	(6, 320)
Fc1	1200	Unknown	512
Fc2	4000	Unknown	1024
Output Size	10 or 40	10 or 40	10 or 40