Free Access

Table 1.

Capsule network architectures trained on the CFHT dataset.

Convolutional layers
Capsule layers (C, D)
Number Feature maps ConvCaps Total
Model F. maps M × (F, N, S) (F, I, D, S) 1 2 3 weights
1 5 3 × (5, 256, 1) (5, 40, 12, 2) (30, 15) (1, 25) 9 861 010
2 × (5, 256, 2)
2 2 2 × (5, 256, 2) (5, 40, 12, 2) (30, 15) (1, 25) 4 945 042
3 2 2 × (5, 256, 2) (5, 40, 12, 2) (1, 15) 4 724 992
4 1 1 × (9, 256, 3) (5, 40, 12, 2) (30, 15) (10, 25) (1, 25) 3 428 222
5 1 1 × (9, 256, 3) (5, 40, 12, 2) (30, 15) (1, 25) 3 320 722
ALED-m 1 1 × (9, 16, 3) (5, 32, 8, 2) (24, 12) (8, 16) (1, 16) 216 608
6 1 1 × (9, 16, 3) (5, 24, 8, 2) (16, 10) (4, 12) (1, 12) 117,280
7 1 1 × (9, 16, 3) (5, 6, 4, 2) (8, 6) (4, 8) (1, 8) 13 880
8 1 1 × (9, 8, 3) (5, 4, 4, 2) (8, 6) (3, 8) (1, 8) 5984
9 1 1 × (9, 8, 3) (5, 2, 4, 2) (6, 4) (2, 6) (1, 6) 2816
10 1 1 × (9, 8, 3) (5, 2, 3, 2) (6, 4) (3, 6) (1, 6) 2546

Notes. Models are defined by the feature map(s), ConvCaps layer, and capsule layer(s), where C is the number of capsules in the layer, D is the number of dimensions per capsule, F is the length of each filter, N is the number of filters used, S is the stride, I is the number of capsule types, and M is the number of feature maps. The ‘Number F. Maps’ column lists the total number of feature maps in each model. The ‘Total Weights’ column lists the total number of trainable weights in each model. All models were trained on 250 images.

Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.

Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.

Initial download of the metrics may take a while.