kornia/tfeat

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

TFeat is a 128-dimensional patch descriptor trained with a triplet loss and a shallow two-branch CNN. One of the first learning-based descriptors to outperform SIFT on standard patch benchmarks. BMVC 2016.

Original repo: vbalnt/tfeat Original license: BSD

Weights

File	Training data
tfeat-liberty.params	Liberty
tfeat-notredame.params	Notre Dame
tfeat-yosemite.params	Yosemite

Citation

@inproceedings{TFeat2016,
    author    = {Vassileios Balntas and Edgar Riba
                 and Daniel Ponsa and Krystian Mikolajczyk},
    title     = {Learning local feature descriptors with triplets
                 and shallow convolutional neural networks},
    booktitle = {BMVC},
    year      = {2016}
}