Warped convolutions: Efficient invariance to spatial transformations
Convolutional Neural Networks (CNNs) are extremely efficient, since they exploit the inherent translation-invariance of natural images. However, translation is just one of a myriad of useful spatial transformations. Can the same efficiency be attained when considering other spatial invariances? Such...
| Main Authors: | , |
|---|---|
| Format: | Conference item |
| Language: | English |
| Published: |
Proceedings of Machine Learning Research
2017
|
| _version_ | 1797078330621558784 |
|---|---|
| author | Henriques, J Vedaldi, A |
| author_facet | Henriques, J Vedaldi, A |
| author_sort | Henriques, J |
| collection | OXFORD |
| description | Convolutional Neural Networks (CNNs) are extremely efficient, since they exploit the inherent translation-invariance of natural images. However, translation is just one of a myriad of useful spatial transformations. Can the same efficiency be attained when considering other spatial invariances? Such generalized convolutions have been considered in the past, but at a high computational cost. We present a construction that is simple and exact, yet has the same computational complexity that standard convolutions enjoy. It consists of a constant image warp followed by a simple convolution, which are standard blocks in deep learning toolboxes. With a carefully crafted warp, the resulting architecture can be made equivariant to a wide range of two-parameter spatial transformations. We show encouraging results in realistic scenarios, including the estimation of vehicle poses in the Google Earth dataset (rotation and scale), and face poses in Annotated Facial Landmarks in the Wild (3D rotations under perspective). |
| first_indexed | 2024-03-07T00:30:33Z |
| format | Conference item |
| id | oxford-uuid:7fa292e7-9e11-4df0-be7d-06010f52b900 |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-03-07T00:30:33Z |
| publishDate | 2017 |
| publisher | Proceedings of Machine Learning Research |
| record_format | dspace |
| spelling | oxford-uuid:7fa292e7-9e11-4df0-be7d-06010f52b9002022-03-26T21:18:11ZWarped convolutions: Efficient invariance to spatial transformationsConference itemhttp://purl.org/coar/resource_type/c_5794uuid:7fa292e7-9e11-4df0-be7d-06010f52b900EnglishSymplectic Elements at OxfordProceedings of Machine Learning Research2017Henriques, JVedaldi, AConvolutional Neural Networks (CNNs) are extremely efficient, since they exploit the inherent translation-invariance of natural images. However, translation is just one of a myriad of useful spatial transformations. Can the same efficiency be attained when considering other spatial invariances? Such generalized convolutions have been considered in the past, but at a high computational cost. We present a construction that is simple and exact, yet has the same computational complexity that standard convolutions enjoy. It consists of a constant image warp followed by a simple convolution, which are standard blocks in deep learning toolboxes. With a carefully crafted warp, the resulting architecture can be made equivariant to a wide range of two-parameter spatial transformations. We show encouraging results in realistic scenarios, including the estimation of vehicle poses in the Google Earth dataset (rotation and scale), and face poses in Annotated Facial Landmarks in the Wild (3D rotations under perspective). |
| spellingShingle | Henriques, J Vedaldi, A Warped convolutions: Efficient invariance to spatial transformations |
| title | Warped convolutions: Efficient invariance to spatial transformations |
| title_full | Warped convolutions: Efficient invariance to spatial transformations |
| title_fullStr | Warped convolutions: Efficient invariance to spatial transformations |
| title_full_unstemmed | Warped convolutions: Efficient invariance to spatial transformations |
| title_short | Warped convolutions: Efficient invariance to spatial transformations |
| title_sort | warped convolutions efficient invariance to spatial transformations |
| work_keys_str_mv | AT henriquesj warpedconvolutionsefficientinvariancetospatialtransformations AT vedaldia warpedconvolutionsefficientinvariancetospatialtransformations |