---
license: mit
tags:
  - kornia
  - feature-description
  - patch-descriptor
---

# kornia/hardnet

Pretrained weights for **HardNet** and **HardNet++**,
used by [`kornia.feature.HardNet`](https://kornia.readthedocs.io/en/latest/feature.html).

HardNet is a 128-dimensional descriptor for 32×32 grayscale patches, trained with a
hard-negative mining triplet loss. NeurIPS 2017.

**Original repo:** [DagnyT/hardnet](https://github.com/DagnyT/hardnet)

## Weights

| File | Training data |
|------|--------------|
| `HardNetPP.pth` | All Brown datasets (HardNet++) |
| `checkpoint_liberty_with_aug.pth` | Liberty with photometric augmentation |
| `hardnet8v2.pt` | HardNet8 v2, trained on HPatches |

## Citation

```bibtex
@inproceedings{HardNet2017,
    title     = {Working hard to know your neighbor's margins:
                 Local descriptor learning loss},
    author    = {Anastasiya Mishchuk and Dmytro Mishkin
                 and Filip Radenovic and Jiri Matas},
    booktitle = {NeurIPS},
    year      = {2017}
}

@article{HardNet2020,
    title   = {Improving the HardNet Descriptor},
    author  = {Milan Pultar},
    journal = {arXiv ePrint 2007.09699},
    year    = {2020}
}
```