Benchmark: Multi-level Dense Capsule Networks

[ssrp]

The model details are as follows:
Architecture is an improvement of the baseline capsule networks, proposed by Sabour et al., NIPS 2017. Accepted in ACCV 2018: Code and data could be found in the GitHub Repository.
Preprocessing: Normalization and shift at most 2 pixel and horizontal flip
Accuracy on MNIST: 99.75%
Accuracy on F-MNIST: 94.65%