No description, website, or topics provided.
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
evaluation
examples/car
include
src
tools
README.md
caffe.proto

README.md

Code for AAAI 2019 Paper "Energy Confused Adversarial Metric Learning for Zero-Shot Image Retrieval and Clustering"

This code is developed based on Caffe.

This code also consists the baseline methods used in our paper, such as Triplet loss, Npair loss and Binomial loss.

Prerequisites

  • Caffe
  • GPU Memory >= 11G
  • Matlab(used for evaluation)

Training

  • Add our files (in include/, src/ and tools/) to your caffe master
  • Merge file caffe.proto into your own caffe.proto
Note only that you should replace your code of {message InnerProductParameter{...}} with our provided code
  • Then complie caffe
 cd ~/caffe-master 
 make all -j8
  • Copy our examples/car/ to your caffe-master, and create folders run/, features/, car_ims_256_nopadding/
  cp -r ~/ECAML-master/examples/car ~/caffe-master/examples/
  mkdir ~/caffe-master/examples/car/run
  mkdir ~/caffe-master/examples/car/features
  mkdir ~/caffe-master/examples/car/car_ims_256_nopadding
  • Download images of Cars196 dataset, and move car_ims/ to ~/caffe-master/examples/car/car_ims_256_nopadding
  • Download our training_list (~200MB), and move it to ~/caffe-master/examples/car/
you can create your own list, as in our paper, the list is created by random selecting in 65*2 manner (with 65 classes and 2 images per class)
  • Download googlenet model to ~/caffe-master/examples/car/
  • Then train the model by ./finetuen.sh
uncomment "energy confusion" codes in train_confusion.prototxt to use our method

Extract features

  • After training, the model will be stored at folder run/, then run ./extractfeatures.sh to extract testing image features(feature files will be stored in folder features/)

Evaluation

  • run code in folder ~/ECAML-master/evaluation

Citation

If our code is helpful, please kindly cite the following papers:

@inproceedings{songCVPR16,
    Author = {Hyun Oh Song and Yu Xiang and Stefanie Jegelka and Silvio Savarese},
    Title = {Deep Metric Learning via Lifted Structured Feature Embedding},
    Booktitle = {Computer Vision and Pattern Recognition (CVPR)},
    Year = {2016}
}
@InProceedings{chen2019energy,
author = {Chen, Binghui and Deng, Weihong},
title = {Energy Confused Adversarial Metric Learning for Zero-Shot Image Retrieval and Clustering},
booktitle = {AAAI Conference on Artificial Intelligence},
year = {2019}
}