In the first set which went online there were some errors. Most notably one subset being included twice. Also some transposed images. Tests on the old set are invalid.
The set consists of N groups of 4 images each. All the images are 640x480.
If you use the dataset, please refer to:
D. Nistér and H. Stewénius. Scalable recognition with a vocabulary tree. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), volume 2, pages 2161-2168, June 2006. [ bib | .ppt | .pdf ]
Subsets
For users of subsets of the database please note that the difficulty is dependent on the chosen subset. Important factors are:
Difficulty of the objects themselves. CD-covers are much easier than flowers. See performance curve below.
Sharpness of the images. Many of the indoor images are somewhat blurry and this can affect some algorithms.
Similar or identical objects. All the pictures where taken by CS students/faculty/staff and thus keyboards and computer equipment are popular motives. So is computer vision literature.
Download
Please note BEFORE starting your download that the file is almost 2GB. Please save a local copy in order to save bandwidth at our server.
Visual Words. We extracted visual words for each document and wrote them one document per line. Data before ":" is header and then data. The vocabulary was 6 levels and splitting with a factor of 10. The vocabulary was trained on non-related data.
Performance
In the paper we give results either for a subset of 6376 images (all we had at that time) or a smaller subset of 1400 images. The smaller set was used when we did not have an efficient enough implementation in order to handle the larger set.
Performance Measures
Our simplest measure of performance is to count how many of the 4 images which are top-4 when using a query image from that set of four images.
A matlab implementation which computes this measure: Download.
Numbers for computing our measure on the full 10200 database using different training-sets and different scoring strategies:
How our performance varies when taking subsets 0:n from the set. The different curces represent different choices in scoring strategy. For extremely fast applications we use the flat-scoring while for better performance we use hierarchical scoring. The feature extractor was set to use relatively few features for these experiments.
How the score is computed
int nrblocks = nr_docs/4;
int totaltopcount = 0;
for( block = 0; block < nrblocks; block++) {
for( int i=0; i < 4; i++){
int pos = block*4+i;
for( int j=0; j < 4; j++){
r = find_rank_of_doc (4*block+j) relative to doc (block*4+i);
if( r < 4)
totaltopcount++;
}
}
}
score = totaltopcount/(nrblocks*4);
What we are measuring is how many of the images are found on average.
Getting everything right gives a score of 4
Getting nothing right gives a score of 0
Getting only identical image right gives a score of 1
A score of 3 means that we find the identical image plus 2 of the 3 other images of the set.
Semiprocessed Data
We have computed lots of semiprocessed data along with SIFT vectors for training.