A New Evaluation Protocol and Benchmarking Results for Extendable Cross-media Retrieval
Ruoyu Liu, Yao Zhao, Liang Zheng, Shikui Wei, Yi Yang
Cross-media retrieval studies the problem that the modelities of the query and database are not the same. In previous works, the performance of cross-media is tested on the classical datasets and the train/test classes are identical. However, the query image/text query may exhibit various content and it is challenging for the training process to take into all the variety in query types. It is therefore not well-grounded for an evaluaiton protocol to assume that the train/test data have the same set of classes. We propose a new evaluation protocol of the extendable cross-media retrieval and re-evaluate the performance of the baseline methods.
Detailed description is provided in our paper.
This software is both tested on Windows 8.1 and Ubuntu 16.04 LTS (64bit).
MATLAB (tested with 2015a on Windows and 2016b on 64-bit Linux).
Caffe (also provided in our project).
You can directly run the three script files in "crossmedia" to view the results of real-valed representation methods, binary representition methods and trivial solution methods. The results are also saved in this folder.
>> run runrl.m >> run runhs.m >> run runts.m
Train the Deep Models
To train the deep models tested in our paper. You need to download the data of the three benchmark datasets from here (password: ivms), unzip the file and put the inside folders into "Benchmarks". Then, you need to download the pre-trained CaffeNet and the image mean value file from here (password: 8krs). Unzip the file and the put the inside two files into "Crossmedia(deep)".
Modify the data path variables of the script files in "Crossmedia(deep)/Codes", then run these files to generate the training data.
Modify the data path of the ".sh" and ".prototxt" files in "Crossmedia(deep)/Networks", and then run the ".sh" files to train the models and extract the features.
Copy the new feature files (".nn") to the folder "Crossmedia/deep_features".
You can download the result figures illustrated in our paper (saved as ".fig" files) from "Figs" for direct comparison.