Keypoint-Descriptor-Optimization / hpatches-benchmark

Python & Matlab code for local feature descriptor evaluation with the HPatches dataset.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

logo

Homography patches dataset

This repository contains the code for evaluating feature descriptors on the HPatches dataset. For more information on the methods and the evaluation protocols please check [1].

Benchmark implementations

We provide two implementations for computing results on the HPatches dataset, one in python and one in matlab.

python matlab
details details

Benchmark tasks

Details about the benchmarking tasks can he found here.
For a more in-depth description, please see the CVPR 2017 paper [1].

Getting the dataset

The data required for the benchmarks are saved in the ./data folder, and are shared between the two implementations.

To download the HPatches image dataset, run the provided shell script with the hpatches argument.

sh download.sh hpatches

To download the pre-computed files of a baseline descriptor X on the HPatches dataset, run the provided download.sh script with the descr X argument.

To see a list of all the currently available descriptor file results, run scipt with only the descr argument.

sh download.sh descr       # prints all the currently available baseline pre-computed descriptors
sh download.sh descr sift  # downloads the pre-computed descriptors for sift

The HPatches dataset is saved on ./data/hpatches-release and the pre-computed descriptor files are saved on ./data/descriptors.

Dataset description

After download, the folder ../data/hpatches-release contains all the patches from the 116 sequences. The sequence folders are named with the following convention

  • i_X: patches extracted from image sequences with illumination changes
  • v_X: patches extracted from image sequences with viewpoint changes

For each image sequence, we provide a set of reference patches ref.png. For the remaining 5 images in the sequence, we provide three patch sets eK.png and hK.png and tK.png, containing the corresponding patches from ref.png as found in the K-th image with increasing amounts of geometric noise (e<h<t).

patches

Please see the patch extraction method details for more information about the extraction process.

References

[1] HPatches: A benchmark and evaluation of handcrafted and learned local descriptors, Vassileios Balntas*, Karel Lenc*, Andrea Vedaldi and Krystian Mikolajczyk, CVPR 2017. *Authors contributed equally.

You might also be interested in the 3D reconstruction benchmark by Schönberger et al. also presented at CVPR 2017.

About

Python & Matlab code for local feature descriptor evaluation with the HPatches dataset.

License:BSD 2-Clause "Simplified" License


Languages

Language:MATLAB 65.7%Language:Python 33.1%Language:Shell 1.2%