| Author | Denis Gudkov |
| Consultant | Daniil Dorin |
| Advisor | Andrii Hrabovyi |
A pirated image is rarely a pixel-perfect copy: rotations, reflections, crops, and re-compression are routine, and a detector should map all of them to the same point in feature space. We ask whether building these symmetries into the encoder of a siamese plagiarism detector is preferable to approximating them with data augmentation, and we study this question separately for the dihedral group of the square and for the special Euclidean group of the plane. Three encoders are compared inside one pipeline: a non-equivariant ViT-Tiny baseline trained with group-specific geometric augmentation, an Octic ViT with exact dihedral invariance, and a Harmformer with continuous roto-translational invariance. Training pairs are formed on COCO; evaluation uses cross-domain pairs from DomainNet, both overall and split by one geometric transformation at a time. For the dihedral group, the architecturally invariant Octic ViT trained without geometric augmentation matches the augmented baseline on every dihedral transformation and improves precision on negatives, which supports the equivariance hypothesis. For the special Euclidean group, the verdict is mixed: the Harmformer is exact on right-angle rotations and reflections, but collapses on off-axis rotations and on translation, which we attribute to interpolation noise at low input resolution rather than to the equivariance principle.
If you find our work helpful, please cite us.
@article{citekey,
title={Title},
author={Name Surname, Name Surname (consultant), Name Surname (advisor)},
year={2025}
}Our project is MIT licensed. See LICENSE for details.