Name: crow
Owner: Yahoo Inc.
Description: Cross-dimensional weighting for aggregated deep convolutional features.
Created: 2016-02-24 15:15:18.0
Updated: 2018-05-24 06:06:35.0
Pushed: 2016-02-26 02:41:24.0
Size: 17
Language: Python
GitHub Committers
| User | Most Recent Commit | # Commits |
|---|
Other Committers
| User | Most Recent Commit | # Commits |
|---|
This repository contains a Python implementation of CroW aggregation for deep convolution features. CroW is an efficient non-parametric weighting and aggregation scheme to transform convolutional image features to a compact global image feature. This repository contains code to evaluate these global CroW features for image retrieval on common retrieval benchmarks. A full description of this work can be found at http://arxiv.org/abs/1512.04065.
This repository contains scripts to download the Oxford and Paris image retrieval benchmark datasets, as well as the VGG16 pre-trained convolutional neural network model for feature extraction. The Python scripts in this repository offer general utilities for extracting features from Caffe models, experimenting with different feature aggregation schemes, and evaluating retrieval performance on the Oxford and Paris datasets. The model and dataset requirements are only necessary to run evaluation.
For feature extraction, this repository requires a working Caffe installation with pycaffe - more information on installing Caffe can be found here. Other Python dependencies can be installed with pip install -r requirements.txt.
The VGG16 model is used for feature extraction. The model parameters can be downloaded into the vgg folder by running:
vgg/get_vgg.sh
The vgg folder contains a version of the VGG16 model prototxt file that is modified to remove all fully-connected layers. This is necessary in order to apply the model fully convolutionally. It is modified from the original version hosted here: http://www.robots.ox.ac.uk/~vgg/research/very_deep/.
The Oxford dataset can be downloaded to the oxford folder by running:
oxford/get_oxford.sh
The Paris dataset can be downloaded to the paris folder by running:
paris/get_paris.sh
The official C++ program provided by the Oxford group, compute_ap.cpp, for computing the mean average precision (mAP) on the retrieval benchmark is provided in this repository for convenience. It is modified to add an explicit include so that it can be compiled everywhere.
-O compute_ap.cpp -o compute_ap
Scripts are included to extract features for a directory of images, crop and extract features for query images in the retrieval datasets, and to evaluate aggregation schemes on the retrieval datasets. The scripts are designed to save raw features to disk after extraction. The evaluation script reads these raw features, applies an aggregation function, computes whitening/PCA, and runs the benchmark evalution. The aggregated features and the whitening/PCA params are not saved by default.
The following extracts features for both retrieval datasets. The extract_queries.py dataset is designed to read and crop the query images in the retrieval datasets. The extract_features.py script is general and can be used to extract features for a directory of images. Please refer to the command line arguments in the script for usage details. The following commands may take a while to run for all images in the dataset.
on extract_features.py --images oxford/data/* --out oxford/pool5
on extract_queries.py --dataset oxford
on extract_features.py --images paris/data/* --out paris/pool5
on extract_queries.py --dataset paris
Running python evaluate.py will evaluation mAP with the default settings defined in the script. Please refer to the command line arguments for options. Also of note is the run_eval function contained in the script. It can easily be used to run many evaluations or provide custom aggregation functions for experimentation.
Code licensed under the Apache License, Version 2.0 license. See LICENSE file for terms.