Our development team has been working on a traffic security management project for a large corporate customer for the past 7 years. The main task was to determine the type of vehicle, assess the behaviour on the road and identify registration plates.
To obtain statistically reliable data, we used the analysis of large volumes of data and worked out algorithms for teaching AI.
After thousands of hours of training and correction, we came to an unobvious conclusion: giving data for training in the format of a video stream is not as effective as teaching using individual frames, that is, the video stream was divided into separate frames and such a breakdown was given for AI processing.
In addition, this approach made it possible to more accurately correct analysis errors - milliards of frames gave at the insistence of stunning data for analysis and date mining received an analysis vector and the basis of corrective patterns. Very unexpected results were given by the technology of attaching implicit features to the data for the AI trainee, so the accuracy of determining the brand of the car increased, if the algorithm was "told" - "the car you found in the luxury configuration" increased.
The accuracy of the determination increased from 77.5% to 82.22%, we found more than one hundred such approaches and technological solutions.
Of course, we can not talk about all the know-how, but our achievements have been used in security and traffic management systems in a number of large megacities for more than three years.
Working on our project, we began to look for areas where our ideas could be as useful as for forecasting, managing and ensuring traffic security in megacities.
Like most enthusiasts of the digital future, we have seen how the cryptocurrency industry is developing, but the most interesting for us was of course NFT! Digital patterns are our passion, and there is such diversity in digital art!
We applied the experience gained to teaching AI and "fed" hundreds of thousands of examples of digital art in order to highlight the most popular and underrated works from the whole set.
In this approach, the experience of adding non-obvious data to the training was very useful for us - we showed the selling price of each sample!
And imagine our surprise when our system was able to predict which NFT from the yet unsold collections would be popular and in demand.
Buying even promising collections on NFT marketplaces, as we have determined, is 45 - 50% less profitable than buying collections directly from authors and placing them on the marketplaces we have already chosen.
In addition, it turned out that if we add a marker of the site where the collection was sold to the training algorithm, then our algorithm, in addition to determining promising NPTs for purchase, is able to tell on which marketplace it is necessary to place the collection to increase the income from the transaction.