Download
Image Database
whether the images contained one salient foreground object. The purpose
is the image from this set can be properly processed by segmentation-based
image retrieval systems. The category we selected from COREL is 290,
700, 750, 770, 840, 1040, 1050, 1070, 1080, 1090, 1100, 1120, 1340, 1350,
1680, 2680, 2890, 3260, 3510, 3540, 3910, 4150, 4470, 4580, 4990, 5210,
5350, 5530, 5810, 5910, 6440, 6550, 6610, 6840. The image is numbered
from 00 to 99. Therefore, the first image's ID in category 290 is 29000
and the last one's in the same category is 29099.
We resized the image so that max(width, height)=384 and min(width, height)=256. The processed image set can be acquired here.
Download: Resized image database, Thumb nail image database (for display).
CBIR Channels
Four image representations are used to generate different CBIR channels,
which are
color (C+), color negative (C-), grayscale (B+), and grayscale-negative
(B-).
Two typical CBIR systems are employed to generate channels. One is global-feature-based and the other is region-based. The globalfeature-based system is named Basic CBIR system and it is built similar to MiAlbum system (Liu 2001) with seven visual features, including three color features and four texture features. All these features are compared respectively to get a similarity and then linearly combined with equal weights. The region-based system employed in this experiment is SIMPLIcity (Wang 2001). This system performs image segmentation and then uses an integrated-region-matching (IRM) approach in distance calculation.
The channel for a given query (an image ID), representation, and retrieval system can be get from file: retrieval_system_name/representation_name/queryID. The channel file is a 2-d table which has 3400 rows and 3 columns.
The format for Basic CBIR channel is listed as:
#rank (from 0-3399) image-ID similarity
(from 0-1)
The format for SIMPLIcity channel is listed as:
#rank (from 0-3399) image-ID distance
(from 0-)
The rows reflect the ranking toward all images given the query. There could be ties in the similarity/distance field.
We also provide a java program to read the CBIR channel. It could read a specific channel file and turn it into a RankList object of our fusion package.
Download: CBIR Channels
Feature Vectors
If you are interested in how our BASIC CBIR system works, here are the features we used. The features can be divided into two main categories: color features and texture features. Table 1 shows the classification of the features in detail. We can find these are actually the classic and typical visual features for a CBIR system. The description for these features can be easily get in literature.
|
Feature Categories |
Feature Name |
Feature Length |
|
|
Color Features |
Color Histogram |
||
|
ColorHsvHistogram64 |
64 |
||
|
Color Moment |
|||
|
ColorLuvMoment123 |
9 |
||
|
Color Coherence |
ColorHsvCoherence64 |
128 |
|
|
Texture Features |
Tamura Texture |
||
|
CoarsnessVector |
10 |
||
|
Directionality |
8 |
||
|
Wavelet Texture |
|||
|
WaveletTwtTexture |
104 |
||
|
MASAR Texture |
MRSAR |
15 |
|
Table1 The classification of the features
These features are compared separately, and then combined linearly by equal weights (they are treated as the same important). Although it is possible to assign different weight for features, it turns out matter little in our experiments. So, we use this method to perform nearest neighbor search in our BASIC CBIR system.
The feature for a given image of one representation can be get from file: representation_name/imageID. The feature file is a plain text which has 7 rows, each row represent one feature vector. Each row has floating numbers divided by spaces representing the components of the vector. The 7 features are listed in sequential order as: ColorHsvHistogram64, ColorLuvMoment123, ColorHsvCoherence64, CoarsnessVector, CoarsnessVector, Directionality, WaveletTwtTexture, and MRSAR. Each feature vector is normalized to has the same length as 1/7. Although for each feature, there is specific distance measure designed for it (for example, intersection of color histogram), The easiest implementation of a basic CBIR system is just concatenate all the 7 features to one 338 dimensional vector and use the Euclidian distance to evaluate image distances.
Download: Feature Vectors
This has been widely used in CBIR evaluation. All the images inside the same COREL category to be relevant.
The images are manually labeled with its foreground object. There are 72 different labels finally for this image set. And each image could have muliple labels if necessary. Any pair of images will be regarded as relevant if they both contains any label.
The groundtruth is provided in the classic 4 column trec_eval format so that it could be read and processed by trec_eval program. It is listed as:
#Query-image-ID Q0 #Image-ID 1
Download here: COREL groundtruth, Fore groundtruth.