Skip to content

Latest commit

 

History

History
113 lines (85 loc) · 3.65 KB

README.md

File metadata and controls

113 lines (85 loc) · 3.65 KB
Raw

VLB-Deteval

Source code for Large scale evaluation of local image feature detectors on homography datasets (BMVC 2018) [ArXiv].

This project is based on the VLB library and is written in MATLAB. Binary version can be run with the free MATLAB SDK.

For a tutorial how to evaluate a new detector, please see the following instructions.

Update 07/06/19 The provisioned scores were computed with an incorrect homographies for the EdgeFoci dataset. This has been now fixed (vlb-deteval-scores-*-v2.tar.gz). We would like to thank Ignacio Rocco for notifying us about this and providing a fix.

Installation

If you have MATLAB 2017a (older versions not tested), you can use source code directly. Otherwise you can use the binary distribution with the MATLAB Compiler Runtime (MCR).

Setup

To set up the MATLAB environment and to compile the VLB mex files, simply run:

>> de

This also shows the list of available commands.

Provision data files

To download the compact archive with the final results of each detector (800kiB), run:

>> de provision scores-compact

To download all the results data (2.3GiB), e.g. to view the per-image result), run:

>> de provision scores-all

Additionally, you can download all the detected keypoints (573MiB), needed when specifying a new experiment:

>> de provision features

Reproduce published results

To generate the published figures, you can simply run:

>> de provision scores-compact
>> de results expdef/bmvc_results.json

This will create the results figures and a rank table in ./data/results/bmvc_results/. The figures are exported in png and tikz format. Rank table is in LaTex format. The figures should look like the following example:

BMVC results for VGGH

You can also recompute all results when only features are privisioned:

>> !rm -rf ./data/scores/bmvc_results_*
>> de provision features
>> de results expdef/bmvc_results.json

it might take few hours.

Visualising image matches

To visualise the image matches, provision the full scores files (de provision scores-all). Additionally, a dataset images will be downloaded if not present.

Visualise an image pair

To visualise and image pair of a dataset (task), run:

>> de view matchpair <datasetname> <taskid>

For example, calling view matchpair vggh 1 results in:

Match pair

Visualise detected keypoints

To visualise detections, run:

>> de view detections <datasetname> <featsname> <imid>

this assumes that the features of featsname are provisioned in ./data/features/featsname.

For example, calling view detections vggh m-surf-ms 1 results in:

Detections

Visualise matching results

To visualise matching results, run:

>> de view matches <benchmarkname> <datasetname> <featsname> <taskid>

this assumes that all scores are either provisioned or computed. The benchmark name is for example bmvc_results_N which is the repeatability for the top-N features.

For example, calling view matches bmvc_results_1000 vggh m-surf-ms 1 results in:

Matching results

Authors

  • Karel Lenc - Initial work - lenck

Citation

Please cite us if you use this code:

@article{VlbDet18,
 author = {Karel Lenc and Andrea Vedaldi},
    title = "{Large scale evaluation of local image feature detectors on homography datasets}",
    journal = {BMVC},
    year = 2018,
    month = sept
}