PlaceRecognitionDB is a tool for creating optimally sized databases (containing the minimum number of frames covering the scene) for place recognition task from RGBD and LiDAR data.
The tool supports several basic methods for creating databases, as well as the DominatingSet method for creating optimal databases. A detailed description of all methods for creating databases can be found in our paper.
The tool also contains a global localization pipeline with CosPlace and NetVLAD. The models of these neural networks can be fine-tuned for a particular previously created database. The results of global localization can also be improved with SuperGlue. More details about the available configurations for global localization and the results are available in the paper.
We have also developed a set of metrics that can be used to evaluate the quality of created databases and the accuracy of VPR systems.
For more, please visit the PlaceRecognitionDB documentation. You can also find full information about our research on the website.
You can find detailed information about the data format used in the tool here.
Please check example.ipynb with a small example on creating a database.
This project is licensed under the Apache License — see the LICENSE file for details.