Caffe based image hashing on imagenet dataset. Originally code-named "ImageNetRetrieval". As course project for Advanced Data Structure.
This README.md documents contains a manifest of files in this repository, as well as a guide to get our code running. For detailed description of our work, please refer to our documentation.
- a modified and customized copy of Caffe in
DSH/ - MATLAB scripts for evaluation and visualization in
matlab/ - Fine tuning configurations along with some Shell scripts in
ImageNet/ - CaffeNet binaries and some extra data in
ExData/ - some Python helper scripts to generate lists of filenames
- MATLAB 2015a or newer
- Python 2.7 interpreter
- Ubuntu 14.04 or higher, or other *nix environment with GNU tool chain
- Requirements of Caffe
- Active Internet connection
You need to build our modified version of Caffe, probably with make.
Depending on your environment (whether or not you have a powerful GPU, enough electricity and enough time), modifications on options concerning mode of CPU of GPU may be done in DSH/tools/extract_features_binary.cpp:57, as well as in some *.prototxt files, by the way.
Modifications of absolute paths in ImageNet/*.prototxt files have to be done according to location of the model, dataset and queries. Manifest files to contain all the filenames of dataset and queries respectively have to be generated, with Python scripts getListWithID.py and getListWithLabel.py.
We are not that crazy to train an ImageNet model by ourselves. Please follow instructions from Caffe to download it.
If you are confident about your GPU's capacity and are determined to contribute to global warming, make sure everything is set in ImageNet/finetune_solver.prototxt, modify paths in ImageNet/finetune_caffenet.sh and then execute it.
Make sure everything is set in ImageNet/finetune_net_im.prototxt, modify paths in ImageNet/extract_code.sh and then execute it. Parameters of the command to be executed in this Shell script may have to be modified according to actual size of the query set.
It's time for MATLAB to enter on the stage.
Put the train-code train-list, query-code query-list in folder matlab
modify the virables refering the inputs in solve.m
trainBinaryFile = 'code-5613.dat';
testBinaryFile = 'code-test.dat';
trainListFile = 'list5613.txt';
testListFile = 'test.txt';
then run matlab at the matlab folder
run solve.m
if you want to visualize the result uncomment the code
% compare(resultNFined, resultFined, trainL, testL);
if you want to visualzie, make sure your pictures are saved in the root folder of project.
and funcion
% mapVec = mapTestOnTrain(resultM, trainY, testY, [10:10:500]);
is used to calculate MAP(mean average precision)
the last parameter is the n-top vector
####Visualizaion sample
