Downloaded 356 times
![Max Jaderberg, Karen Simonyan, Andrew Zisserman, Koray Kavukcuoglu
[arXiv] [GitXiv] [slides] [video]
Spatial transformer
networks
Slides by Víctor Campos [GDoc]
Computer Vision Reading Group (28/03/2016)](https://image.slidesharecdn.com/06spatialtransformernetworks-160329154407/85/Spatial-transformer-networks-1-320.jpg?cb=1459266300)
Uploaded byUniversitat Politècnica de Catalunya
Spatial transformer networks
Presentation by Victor Campos at the Universitat Politecnica de Catalunya (UPC) of the paper: Jaderberg, Max, Karen Simonyan, and Andrew Zisserman. "Spatial transformer networks." In Advances in Neural Information Processing Systems, pp. 2008-2016. 2015. Part of the Computer Vision Reading Group. Full list of papers and schedules: https://github.com/imatge-upc/readcv
Spatial transformer networks
- 1. Max Jaderberg, KarenSimonyan, Andrew Zisserman, Koray Kavukcuoglu [arXiv] [GitXiv] [slides] [video] Spatial transformer networks Slides by Víctor Campos [GDoc] Computer Vision Reading Group (28/03/2016)
- 2. 1. Introduction Why do weneed spatial transformer networks?
- 3. Why do weneed Spatial transformer networks? Are Convolutional Neural Networks invariant to… ▪ Scale? ▪ Rotation? ▪ Translation?
- 4. Why do weneed Spatial transformer networks? CS231n: Convolutional Neural Networks for Visual Recognition (Stanford)
- 5. Why do weneed Spatial transformer networks? Are Convolutional Neural Networks invariant to… ▪ Scale? No ▪ Rotation? ▪ Translation?
- 6. Why do weneed Spatial transformer networks? Are Convolutional Neural Networks invariant to… ▪ Scale? No ▪ Rotation? No ▪ Translation?
- 7. Why do weneed Spatial transformer networks? Are Convolutional Neural Networks invariant to… ▪ Scale? No ▪ Rotation? No ▪ Translation? Partially
- 8. Why do weneed Spatial transformer networks? A. W. Harley, "An Interactive Node-Link Visualization of Convolutional Neural Networks," in ISVC, pages 867-877, 2015
- 9.
- 10.
- 11.
- 12.
- 13. Formulating Spatial transformers Three maindifferentiable blocks: ▪ Localisation network ▪ Grid generator ▪ Sampler
- 14. Grid generator: Examples Affine transformAttention model - coordinates in the target (output) feature map - coordinates in the source (input) feature map
- 15. Sampler: Mathematical formulation Generic sampling kernel Fromthe grid generator All pixels in the output feature map
- 16.
- 17.
- 18. Experiment #1: Distorted MNIST Distortions:Rotation, Translation, Projective, Elastic Transformations: Affine, Projective, Thin Plate Spline (TPS)
- 19. Experiment #2: Multiple spatialtransformers
- 20. Experiment #2: Adding twodigits in an image
- 21. Experiment #2: Adding twodigits in an image
- 22. Experiment #2: Adding twodigits in an image
- 23. Other experiments: Applications ofspatial transformers ▪ Street View House Numbers ▪ Fine-grained classification
- 24.
- 25. Conclusions: Spatial transformer networks ▪A module that performs spatial transformations to features has been presented ▪ It is differentiable and learnt in an end-to-end fashion ▪ No modifications to the loss function are needed ▪ Outperforms the state-of-the-art performance in some tasks
- 27.