kornia/affnet

Pretrained weights for AffNet, used by kornia.feature.LAFAffineShapeEstimator.

AffNet is a CNN that estimates the 2×2 affine shape matrix for a detected keypoint patch, making subsequent description invariant to affine transformations. Trained jointly with HardNet to maximise discriminability of affine-normalised patches. ECCV 2018.

Original repo: ducha-aiki/affnet

Weights

File	Description
AffNet.pth	Default checkpoint

Citation

@inproceedings{AffNet2018,
    author    = {D. Mishkin and F. Radenovic and J. Matas},
    title     = {{Repeatability is Not Enough}: Learning Affine Regions
                 via Discriminability},
    booktitle = {ECCV},
    year      = {2018}
}