libvot is a fast implementation of vocabulary tree, which is an algorithm widely used in image retrieval and computer vision. It usually comprises three components to build a image retrieval system using vocabulary tree: build a k-means tree using sift descriptors from images, register images into the database, query images against the database. In this library, we use C++11 standard multi-thread library to accelerate the computation, which achieves fast and accurate image retrieval performance. Currently this library is under active development for both research and production. If you find this repository useful, please star it to let me know. :)

The build system of libvot is based on CMake. To take full advantages of the new features in C++11, we require the version of CMake to be 2.8.12 or above. Current we have tested our project under Linux (Ubuntu 14.04, CentOS 7) and MacOS (10.10) using gcc. The common steps to build the library are:

1. Extract source files.
2. Create build directory and change to it.
3. Run CMake to configure the build tree.
4. Build the software using selected build tool.
5. Run "make test"
6. See src/example for the usage of this library.

On Unix-like systems with GNU Make as the build tool, the following sequence of commands can be used to compile the source code.

$ cd libvot
$ mkdir build && cd build
$ cmake ..
$ make && make test

• gflags, glog (Required) and gtest, if test is enable.
• Boost (>1.55), for serialization, python-binding, etc.
• OpenCV (>2.4), for feature detector and general utilities for image processing.
• NVIDIA's Cuda Toolkit 7.5, for GPU-related code.
• NVIDIA's cuDNNv5 for CUDA 7.5, for the deep learning module.

See the installation guide for details.

Besides, libvot supports docker installation. Docker is a system to build self-contained versions of a Linux operating system running on your machine. You can pull the latest auto-build here. If you encounter any problem building this software on a clean linux OS, Dockerfile is a minimum ubuntu configuration and a good reference.

Suppose $LIBVOT_ROOT represents the root directory of libvot, and it is successfully compiled in build subdirectory. You can use ./libvot_feature <image_list> to first generate a set of descriptor files and use them as inputs to image_search. For example, you have some target .jpg image files to generate sift files. Just cd into that directory, prepare the image_list, and generate sift files in the same directory as the image files:

$ ls -d $PWD/*.jpg > image_list
$ $LIBVOT_ROOT/build/bin/libvot_feature <image_list>

Then you can run image_search in src/example to generate the image retrieval results The usage is simply “./image_search <sift_list> <output_dir> [depth] [branch_num] [sift_type] [num_matches] [thread_num]”. We add a small image dataset fountain-P11 to illustrate this process. test_data/fountain_dense folder contains the sift files generated by libvot_feature, while the original images are not included in order to save space. If you use the out-of-source build as shown in the installation section and in the build directory, the following command should work smoothly and generate several output files in build/bin/vocab_out directory (you need to change the prefix of filepaths in test_data/fountain_dense/sift_list).

$ cd bin
$ ls -d $PWD/*.sift > sift_list
$ ./image_search <sift_list> <output_folder> [depth] [branch_num] [sift_type] [num_matches] [thread_num]  
$ (e.g.) ./image_search ../../test_data/fountain_dense/sift_list ./vocab_out

Each line in match.out contains three numbers “first_index second_index similarity score”. Since the library is multi-threaded, the rank is unordered with respect to the first index (they are ordered w.r.t the second index). match_pairs saves the ordered similarity ranks, from 0th image to n-1th image. Besides, libvot also supports sift files generated by openMVG (set [sift_type] to 1).

The homepage of libvot is hosted by github-pages. See the documentation here.

We are working toward the next major release (0.2.0). If you are interested in contributing, please have a look at Roadmap.md and our Coding style. All types of contributions, including documentation, testing, and new features are welcomed and appreciated.

If you find this library useful for your research, please cite

@inproceedings{shen2016graph,  
  title={Graph-Based Consistent Matching for Structure-from-Motion},  
  author={Shen, Tianwei and Zhu, Siyu and Fang, Tian and Zhang, Runze and Quan, Long},  
  booktitle={European Conference on Computer Vision},  
  pages={139--155},  
  year={2016},  
  organization={Springer}  
}

Note: The image retrieval part of the above research depends on libvot. The functioning graph matching algorithm is in preparation and is planned to be merged into the master branch. For an early preview and implementation details, please send your request to [email protected].

The BSD 3-Clause License

For inquiries and suggestions, please send your emails to [email protected].

If you would like to support this project, you can contribute to this project, or make a donation via pledgie. Thanks!