Recognizing objects
using a billion-image
database

The FindFace database stores image vectors—basically arrays of numbers. At the input, the system accepts a video stream from security cameras. The system recognizes target images such as faces, silhouettes, car models, and license plates. It also records new vectors. These vectors are then compared with the number arrays that are already in the database. Finally, the system displays search results.

In 2016, FindFace was a working prototype that encountered problems like database corruption and lack of scalability. At the same time, new functionality and reliability requirements arose. To make FindFace a scalable out-of-the-box solution, the developers had to do the following:

1. Increase operation speed. Running pairwise comparisons on objects from large databases took more than 10 seconds and put a burden on the DBMS. Even with the use of indexes, it was impossible to avoid comparisons with all the images in the database.
2. Make database expansion possible. Not only image vectors, but event metadata (like date of detection or camera name) was highly valuable. This metadata is necessary for queries like «show the top 10 most similar faces that were on cameras 1 and 2 from Thursday to Friday». It was kept apart from the vectors, which was inconvenient. It was impossible to filter metadata by index quickly, and sometimes, there were desynchronization issues.

Results

The FindFace service can process information from cameras in less than 0.5 seconds, even if the database contains millions of images. A full walk through the database in search of a similar image no longer overloads the DBMS and is now six times faster. Note that Tarantool incorporates several functions: saving images to the database, searching through the databases, and implementing direct queries to the databases.