#content table.checker { border-collapse: collapse ; margin-left: auto ; margin-right: auto ; } #content table.checker td { background-color: #f6f6f6 ; border: 1px solid #DDD ; padding: 0.5em ; }

This page lists a number of example VLFeat applications. The code can be found in the VLROOT/apps/ subdirectory in the VLFeat package.

Basic recognition

Caltech-101 Collage

This sample application uses VLFeat to train an test an image classifier on the Caltech-101 data. The classifier achieves 65% average accuracy by using a single feature and 15 training images per class. It uses:

The program is fully contained in a single MATLAB M-file, and can also be simply adapted to use your own data (change conf.calDir).

Advanced encodings for recognition

This example application extends the Caltech-101 demo above in many ways: it supports multiple encoding methods, including BoVW, VLAD, and Fisher Vectors, tweaked image features, and multiple benchmark datasets. The code is located int apps/recognition. Start from the main file.

The following tables report results on a few standard benchmark datasets (PASCAL VOC 2007 classification challenge, Caltech 101 30 training images, MIT Scene 67, and Flickr Material Dataset) for a number of different encodings:

methodVOC07Caltech 101Scene 67FMD
FV59.12% mAP73.02% Acc58.25% Acc59.60% Acc
FV + aug.60.25% mAP75.61% Acc57.57% Acc60.80% Acc
FV + s.p.62.23% mAP77.63% Acc61.83% Acc60.80% Acc
VLAD + aug.54.66% mAP78.68% Acc53.29% Acc49.40% Acc
BOVW + aug.49.87% mAP75.98% Acc50.22% Acc46.00% Acc

The baseline feature is SIFT (vl_dsift) computed at seven scales with a factor $\sqrt{2}$ between successive scales, bins 8 pixel wide, and computed with a step of 4 pixels. All experiments but the Caltech-101 ones start by doubling the resolution of the input image. The details of the encodings are as follows: