Skip to content
プロジェクト

DagnyT/hardnet has 535 GitHub stars and 101 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-19 04:34 UTC. Last push 2025-04-14.

この数値の測り方

計測したスター増加

24時間
0
7日間
履歴を収集中
15日間
履歴を収集中
1か月
履歴を収集中

期間ごとの純増数です。期間は重複しており、日別の値ではありません。未計測の値は空欄です。スター数は注目度の指標であり、品質の保証ではありません。

概要

Hardnet descriptor model - "Working hard to know your neighbor's margins: Local descriptor learning loss"

機械学習

HardNet model implementation

HardNet model implementation in PyTorch for NIPS 2017 paper "Working hard to know your neighbor's margins: Local descriptor learning loss" poster, slides

An example how to compile HardNet to Torchscript to be used in C++ code

Notebook

Update April 2025

The PhotoTour dataset is no longer available at original links. We use a mirror from CTU in Prague instead.

Update April 06 2018

We have added small shift and rot augmentation, which improves results up to 1mAP point on HPatches. It is in HardNet.py, turn on by --augmentation=True. All the weight will be updated soon. Version, which is trained on Brown + HPatches + PS datasets is in progress, stay tuned :)

Re: popular question about BoW retrieval engine

Unfortunately, it is proprietary and we cannot release it. But you can try the following open source repos, both Matlab-based:

Benchmark on HPatches, mAP

HPatches-results

Retrieval on Oxford5k, mAP, Hessian-Affine detector

Descriptor BoW BoW + SV BoW + SV + QE HQE + MA
TFeatLib 46.7 55.6 72.2 n/a
RootSIFT 55.1 63.0 78.4 88.0
L2NetLib+ 59.8 67.7 80.4 n/a
HardNetLibNIPS+ 59.8 68.6 83.0 88.2
HardNet++ 60.8 69.6 84.5 88.3
HesAffNet + HardNet++ 68.3 77.8 89.0 89.5

Requirements

Please use Python 2.7, install OpenCV and additional libraries from requirements.txt

Datasets and Training

To download datasets and start learning descriptor:

git clone https://github.com/DagnyT/hardnet
./code/run_me.sh

Logs are stored in tensorboard format in directory logs/

Pre-trained models

Pre-trained models can be found in folder pretrained.

3rd party pre-trained models

Rahul Mitra presented new large-scale patch PS-dataset and trained even better HardNet on it. Original weights in torch format are here.

Converted PyTorch version is here.

HardNet-Datasets-results

Which weights should I use?

For practical applications, we recommend HardNet++.

For comparison with other descriptors, which are trained on Liberty Brown dataset, we recommend HardNetLib+.

For the best descriptor, which is NOT trained on HPatches dataset, we recommend model by Mitra et.al., link in section above.

Usage example

We provide an example, how to describe patches with HardNet. Script expects patches in HPatches format, i.e. grayscale image with w = patch_size and h = n_patches * patch_size

cd examples
python extract_hardnet_desc_from_hpatches_file.py imgs/ref.png out.txt

or with Caffe:

cd examples/caffe
python extract_hardnetCaffe_desc_from_hpatches_file.py ../imgs/ref.png hardnet_caffe.txt

Projects, which use HardNet

AffNet -- learned local affine shape estimator.

Citation

Please cite us if you use this code:

@inproceedings{HardNet,
 author = {Anastasiya Mishchuk, Dmytro Mishkin, Filip Radenovic, Jiri Matas},
 title = "{Working hard to know your neighbor's margins: Local descriptor learning loss}",
 booktitle = {Proceedings of NeurIPS},
 year = 2017,
 month = dec
}