Speaker
Description
Image analysis methods are sensitive to the orientation of the inputs.
This has direct implications on the performance in their respective application fields, like astronomy, bio-medical imaging, as well as many technical tasks.
While in the past, researchers usually focused on data augmentation and brute force approaches, we bring forward a novel idea that utilizes the concept of equivariance in $O(2)$ to achieve better generalization to unseen orientations.
The core concept of our approach is to represent the image data in the space of Zernike polynomials, for which we derive a learnable equivariant feature map.
In this work, we show both theoretically as well as experimentally, that the presented framework achieves near perfect equivariance.
When used for classification tasks, our method performs equally well as other state-of-the-art methods.
| Affiliation of the submitter | HITS |
|---|---|
| Attendance | in-person |
