BoWG

BoWG is an open-source C++ library for loop closure detection and place recognition, serving as the official implementation of our IROS 2025 paper "Bag-of-Word-Groups (BoWG): A Robust and Efficient Loop Closure Detection Method Under Perceptual Aliasing". For a practical demonstration of integrating BoWG into SLAM systems, we provide another repository BoWG_VINS_Loop, which showcases the implementation of BoWG in VINS-Fusion.

Core Features (Robust, Accurate, and Efficient)

• Visual loop closure detection method robust against perceptual aliasing (such as narrow pipes characterized by feature scarcity and repetitiveness)
• Hierarchical word vocabulary tree (offline trained) combined with dynamically established word group table, delivering context-specific representation
• High computational efficiency and excellent scalability comparable to bag-of-binary-words (DBoW series), making it ideal for large-scale and long-term applications
• Four aspects for computing similarity: Word, Word Group, Feature Distribution, and Temporal Consistency
• Dedicated post-verification modules (temporal and geometrical)
• Supports various types of scoring: L1-norm, L2-norm, Chi-Square, KL-divergence, Bhattacharyya coefficient, Dot Product
• Vocabulary training tools for custom applications
• Ready-to-use demo programs for quick testing with image sequences

BoWG demo detector on New College dataset

VINS-Fusion with BoWG on EuRoC MH_03_medium sequence

1. Prerequisites

1.1 Operating Systems

BoWG is compatible with Ubuntu, macOS, and Windows. Ubuntu is recommended, as other systems may require minor modifications. If you encounter any compatibility issues, please create an issue in our repository with your system details and problem description.

1.2 OpenCV

Please follow the official OpenCV Installation guide.

1.3 Boost

BoWG utilizes the boost::dynamic_bitset class.

For Ubuntu:

$ sudo apt-get install libboost-dev

For macOS:

brew install boost

For Windows:

• Download and install Boost.

2. Build BoWG

To set up BoWG, clone the repository:

git clone https://github.com/EdgarFx/BoWG.git

Build the project:

cd BoWG
mkdir build
cd build
cmake ..
make -j4

3. Demo Usage

1. Download the Public Datasets (New College, City Centre, Bicocca) to YOUR_DATASET_FOLDER
2. Modify the dataset path in config/demo.yaml
3. Move to the build directory and run the demo

With GUI interface:

./demo --gui

The GUI displays processing images and corresponding loop closure results. Press 'p' to pause/resume the program.

To adjust processing speed for better visualization, use the --delay parameter (in milliseconds for each image):

./demo --gui --delay 20

For terminal-only execution:

./demo

Performance metrics (precision, recall, true positives, false positives, etc.) can be computed when ground truth data is provided. For the New College sample, ground truth data is included. The ground truth a .txt file with the following format:

95: 1,2  # Image 95 matches with images 1 and 2
96: 1,2,3
97: 2,3,4
98: 3,4,5

For detailed usage information of the demo program:

./demo --help

4. Run with Your Own Dataset

Our demo program is not restricted to the provided sample datasets, as it implements a ready-to-use detector that can process any sequence of images. To use the demo program with your own dataset:

1. Modify image_path in the config file to point to your image directory
2. Set gt_use to 0 in the config file if ground truth data is unavailable
3. Execute the following commend

./demo --config ../config/my_config.yaml --gui

5. Train Your Own Word Vocabulary

Our vocabulary uses a binary format. Source code is located in ./src/Train_Voc/train_voc.cpp, with the executable in ./build/train_voc.

To train your own vocabulary:

cd build
./train_voc <image_folder> <vocabulary_output>

Please specify the complete path for vocabulary output including the file name, e.g., ./my_voc_folder/brief_voc.bin.

6. Integrate BoWG into SLAM Systems

Please check our additional repository BoWG_VINS_Loop, which demonstrates the integration of BoWG in VINS-Fusion. This project serves as a reference for integrating BoWG into other SLAM systems.

7. Acknowledgements

This project builds upon the foundations of: DBoW2, DLib, DLoopDetector and VINS-Mono.

8. License

The source code is released under BSD 3-Clause License.