This project aims to use metric learning to image search reranking.

Pls read the file USAGE.md

All trainig dataset comes from the MSRAMM dataset.

mLMNN2.4

All testing dataset comes form the WebQueries dataset. Each query test data is stored in $ABSPATH/$dataSetName/$featurename/$data$queryno.mat.

Note: the following $ABSPATH means the project directory.

All model files are stored in $ABSPATH/data/model/$featurename/ . There are two kinds of model, that is , normal models and scale models. The normal models are trained from the extracted feature data. The scale models are trained freom the extracted feature normalized data. We adapted common normalized method, which normalize data between [0,1]. And the normalization uses y = (x - x_min) / (x_max - x_min);

Take the 'gist' feature as example. So, all normal models and scale models are stored in $ABSPATH/data/model/gist/ . For the model trained by the data of MSRAMM query_1. The normal model is $ABSPATH/data/model/gist/gist1.mat, and the scale model is $ABSPATH/data/model/gist/gistscale1.mat.

If you want to use a specific model file, for eaxmpale $ABSPATH/data/model/gist/gistscale1.mat, only use load command in matlab. After file loaded, the matlab working memory has a transform matrix, that is, L. Using transform matrix like Y = L * X, here X is dim-by-sampleno. Each coloum of X and Y represents an example image.

• 2014-06-05 by Aborn Jiang (aborn.jiang@foxmail.com)