README.md

VocabTree2, by Noah Snavely ([email protected])

To build the VocabTree2 library, just do "make" from the top-level directory.

This code is distributed under a FreeBSD license. See the source code for details.

Unfortunately, I can't assist much with problems with the code.

If you find this code useful for your project, please cite the following paper:

Sameer Agarwal, Noah Snavely, Ian Simon, Steven M. Seitz and Richard Szeliski. "Building Rome in a Day." In Proc. ICCV 2009.

How to run

This code works with SIFT keys in the format produced by David Lowe's sift binary. There are three tools provided:

• VocabLearn: build a vocabulary tree from a set of key files using k-means
• VocabBuildDB: build an image database from a set of key files and a tree
• VocabMatch: match an image to a database and retrieve a set of top matches

Example usages:

VocabLearn

Usage: VocabLearn list.in depth branching_factor restarts tree.out

- list.in contains a list of key files, one per line, with each key

file in Lowe's format.

- depth -- depth of tree. 0 indicates a flat tree.

- branching_factor -- number of children each non-leaf node.

- restarts -- number of trials in each run of k-means.

- tree.out -- name of output tree.

Example:

Learn a flat vocabulary tree with 500K visual words using the SIFT keys in list.txt

Store the results in tree.500K.out

./VocabLearn/VocabLearn list.txt 0 500000 1 tree.500K.out

VocabBuildDB

Usage: VocabBuildDB list.in tree.in db.out [use_tfidf:1] [normalize:1] [start_id:0] [distance_type:1]

Example:

./VocabBuildDB/VocabBuildDB list.txt tree.500K.out vocab.db

VocabMatch

Usage: VocabMatch db.in list.in query.in num_nbrs matches.out [distance_type:1] [normalize:1]

Example:

./VocabMatch/VocabMatch vocab.db list.txt query.txt 2 matches.txt

The query file is in the same format as the list file, having one SIFT

key file per line, corresponding to the images to query for matching

to the vocabulary database.

The value num_nbrs is the number of top matches to retrieve for each

query image.

The output matches file has three columns:

- the index of the query image to match

- the index of a well-matching image from the database

- the matching score between these images.

For example, if there are 10 images in the database, 3 query images,

and num_nbrs is 2, this file may look as follows:

0 9 0.8004

0 7 0.3138

1 3 0.8017

1 8 0.3250

2 10 0.7933

2 6 0.3145