In this repository, I will continue adding new layers from recent papers, they are tested on Keras-1.2.2.

The layers include:

[1] Spatial Transformer Networks.

https://arxiv.org/abs/1506.02025

This is based on the STN, instead of using transformation matrices to transform, this layer uses the x,y displacements.

[1] Spatial Transformer Networks.

https://arxiv.org/abs/1506.02025

Related works also include:

[2] Deep Feature Flow for Video Recognition.

https://arxiv.org/abs/1611.07715

Seperate the RNN into two parts, convolution + recurrence. This layer can accept input directly from linear convolution.

[1] https://openreview.net/forum?id=rJJRDvcex&noteId=rJJRDvcex

[1] Layer Normalization.

https://arxiv.org/abs/1607.06450

[1] Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network.

https://arxiv.org/abs/1609.05158

[1] Dynamic Filter Networks.

https://arxiv.org/abs/1605.09673

[1] DeepMatching: Deep Convolutional Matching.

http://lear.inrialpes.fr/src/deepmatching/

[2] Fully-Trainable Deep Matching.

http://www.robots.ox.ac.uk/~tvg/publications/2016/FullyTrainableDeepMatching.pdf

[3] Convolutional neural network architecture for geometric matching.

https://arxiv.org/pdf/1703.05593.pdf

Speed up the im2col operations in several layers, e.g. correlation layer, dynamic filter layer, etc.