kornia/hynet

Pretrained weights for HyNet, used by kornia.feature.HyNet.

HyNet is a 128-dimensional patch descriptor that replaces batch normalisation with Filter Response Normalisation and trains with a hybrid similarity measure combining L2 distance and cosine similarity under a triplet margin loss. NeurIPS 2020.

Original repo: yuruntian/HyNet

Weights

File	Training data
HyNet_LIB.pth	Liberty
HyNet_ND.pth	Notre Dame
HyNet_YOS.pth	Yosemite

Citation

@inproceedings{hynet2020,
    author    = {Tian, Yurun and Barroso Laguna, Axel and Ng, Tony
                 and Balntas, Vassileios and Mikolajczyk, Krystian},
    title     = {{HyNet}: Learning Local Descriptor with Hybrid Similarity
                 Measure and Triplet Loss},
    booktitle = {NeurIPS},
    year      = {2020}
}