H-NeXt: The next step towards roto-translation invariant networks

0578508 - ÚTIA 2024 RIV GB eng C - Conference Paper (international conference)
Karella, Tomáš - Šroubek, Filip - Blažek, Jan - Flusser, Jan - Košík, Václav
H-NeXt: The next step towards roto-translation invariant networks.
34th British Machine Vision Conference 2023. Aberdeen: BMVA, 2023, s. 1-14.
[British Machine Vision Conference 2023 /34./. Aberdeen (GB), 20.11.2023-24.11.2023]
R&D Projects: GA ČR GA21-03921S
Institutional support: RVO:67985556
Keywords : H-NeXT * robustness to unseen deformations * parameter-efficient roto-translation invariant network * classification on unaugmented training set
OECD category: Computer hardware and architecture
http://library.utia.cas.cz/separaty/2023/ZOI/karella-0578508.pdf

The widespread popularity of equivariant networks underscores the significance of parameter efficient models and effective use of training data. At a time when robustness to unseen deformations is becoming increasingly important, we present H-NeXt, which bridges the gap between equivariance and invariance. H-NeXt is a parameter-efficient roto-translation invariant network that is trained without a single augmented image in the training set. Our network comprises three components: an equivariant backbone for learning roto-translation independent features, an invariant pooling layer for discarding roto-translation information, and a classification layer. H-NeXt outperforms the state of the art in classification on unaugmented training sets and augmented test sets of MNIST and CIFAR-10
Permanent Link: https://hdl.handle.net/11104/0347650