Switch from standard Convolution to Deformable Convolution

[weiji14]

Towards improving the realism of our predicted DeepBedMap grids by using Deformable Convolutional Layers! Hopefully able to better capture the streaklines and show less pixelated artifacts.

Inspired by the Winning Solution in NTIRE19 Challenges on Video Restoration and Enhancement (CVPR19 Workshops) - Video Restoration with Enhanced Deformable Convolutional Networks https://xinntao.github.io/projects/EDVR

References:

• Chainer's DeformableConvolution2D Link implementation
• Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., & Wei, Y. (2017). Deformable Convolutional Networks. Retrieved from https://arxiv.org/abs/1703.06211v3
• Wang, X., Chan, K. C. K., Yu, K., Dong, C., & Loy, C. C. (2019). EDVR: Video Restoration with Enhanced Deformable Convolutional Networks. Retrieved from https://arxiv.org/abs/1905.02716v1
• Notes on “Deformable Convolutional Networks”
• Type of convolutions: Deformable and Transformable Convolution

[review-notebook-app[bot]]

Check out this pull request on ReviewNB: https://app.reviewnb.com/weiji14/deepbedmap/pull/171

You'll be able to see notebook diffs and discuss changes. Powered by ReviewNB.

[ghost]

DeepCode Report (#12b8ee)

DeepCode analyzed this pull request. There is 1 new info report.