Content Based Image Retrieval System for Inria Holidays Dataset

This is an explanation of the image retrieval system for all sample images on Inria Holidays Dataset.

Dataset Preparation

The dataset has 500 different groups. The first images for each group are used as query images, others are image database. For this reason, two different folders (query_images and data_images) are created under ./dataset/images/ folder.

Feature Extraction

All images from the given in Inria Holidays Dataset are extracted using 1_feature_extraction.py script under ./src folder. This script is used for extraction three different deep learning models which are VGG19, Inception V3 and Resnet50 models. It reads the RGB images from the ./dataset/images folder, saves the features to the ./dataset/features/ folder.

Nearest Neighbor Approach

Using features for each image, the different k numbers are used in the script ./2_KNN.py to obtain top k nearest (related) images from the database. It saves the results under ./results folder.

K-Means

To evaluate the k-means approach, the 4_k-means.py is used for the dataset. The result files are saved to the ./results folder.

Evaluation

Two different evaluation metric is presented in this GitHub folder.

Ranking Average Precision Score for Nearest Neighbor Approach

In order to evaluate the nearest model with ranking average score, 3_ranking_average_score.py script is used under ./src folder.

mAP Evaluation

To obtain the mAP evaluation metric for each result files which is presented in the dataset, ./eval_holidays/run_map_metric.py. the script is used. It utilizes holidays_map.py script to apply the mAP evaluation metric on each result file.