How to teach convolutional networks to perform spatial transformations? Even though deep learning created a class of compelling models, it still suffers in terms of spatial invariance. We will present the learnable module "Spatial Transformer" published in [1], allowing networks to simply perform transformations such as rotation, scale, transpose or cropping. In the second part of the talk, we will talk about CNN allowing spatial transformation designed for subpixel registration [2-3].
JADERBERG, Max, et al. Spatial transformer networks. Advances in neural information processing systems, 2015, 28. (https://arxiv.org/abs/1506.02025v3)
LEBL, Matej, et al. Subpixel precision in registration of multimodal datasets. In: IOP Conference Series: Materials Science and Engineering. IOP Publishing, 2020. p. 012007. (http://www.utia.cas.cz/biblio?pub=0536185)
BLAŽEK, Jan, et al. Improvement of the visibility of concealed features in artwork NIR reflectograms by information separation. Digital Signal Processing, 2017, 60: 140-151. (http://www.utia.cas.cz/biblio?pub=0468355)
Tutorial on Spatial Transformers (https://pytorch.org/tutorials/intermediate/spatial_transformer_tutorial....)
Attachment | Size |
---|---|
CNN meets spatial transformations.pdf | 926.05 KB |
RegistrationByANN.pdf | 158.21 KB |