MSU Researchers Rethink How Computer Vision Algorithms Are Developed
Scientists at the Faculty of Computational Mathematics and Cybernetics at Lomonosov Moscow State University have investigated why neural-network algorithms for tracking objects in video perform with different levels of accuracy.
The researchers analyzed the internal features that a neural network generates at different stages of image processing and proposed a set of specialized measures they call “micrometrics.” They then compared those measures with actual tracking performance. For one of the metrics, the correlation with algorithm performance on individual video sequences reached 0.96 by Pearson’s correlation coefficient.
We are used to judging AI by its end results: Did it recognize a face, track a car or avoid an obstacle? But what happens inside a neural network in the milliseconds between seeing a frame and making a decision? Usually, that process is a black box. Researchers at the Faculty of Computational Mathematics and Cybernetics at Lomonosov Moscow State University have now proposed a way to look inside it, changing how computer vision algorithms themselves can be developed.
Algorithm Anatomy: Micrometrics Instead of Blind Trial and Error
MSU scientists did not build yet another “super-tracker” promising spectacular accuracy. Their contribution is different: They found a way to understand why some neural-network algorithms for tracking objects in video perform reliably while others fail. By analyzing the internal features a model generates as it processes an image, the researchers derived a set of specialized “micrometrics.”
One of these metrics showed a correlation of up to 0.96 with actual tracking quality, measured by Pearson’s correlation coefficient. That means developers no longer have to blindly cycle through architectures and parameters in the hope of finding a winning combination. They now have a mathematical compass that points to the internal properties of a neural network associated with stable object tracking. In effect, this moves AI development away from something resembling “alchemy” and toward precise, explainable engineering.
A Global Shift: From Detection to Meaningful Tracking
The work by the Moscow mathematicians fits neatly into the broader evolution of the field. In 2021-2022, researchers at the Moscow Institute of Physics and Technology focused on 3D tracking and networks such as FMFNet. By 2024-2025, researchers at Bauman Moscow State Technical University were working on accelerating algorithms for real-time operation. Today, the industry faces a new challenge.
Global technology companies are releasing general-purpose models such as SAM 2 that can segment video effectively but can lose track when objects overlap or lighting becomes difficult. Russian researchers are proposing the next step in that evolution: moving beyond the simple fact of detection toward deeper understanding and reliable tracking under the chaotic conditions of the real world.
A Foundation for Sovereign Technology: From Smart Cities to Exports
Reliable tracking is the nervous system of intelligent transportation systems, industrial robots, logistics and smart-city infrastructure. The more reliably an algorithm can keep an object in a video stream, the fewer critical errors occur during automated data processing.
Russia already has a foundation for this work. Companies such as NtechLab have successfully deployed computer vision solutions across dozens of regions and export them to Global South markets. Joint projects to develop domestic hardware and software systems also require a strong mathematical foundation. MSU’s micrometrics could become the hidden engine that improves the competitiveness of Russian IT products while reducing dependence on off-the-shelf solutions from abroad.
From the Lab to Production
The path from an academic paper presented at the Lomonosov Readings to an industry standard is, of course, a long one. The micrometrics still need to be stress-tested across different architectures, under conditions involving object occlusion and abrupt changes in scale and viewpoint. Researchers will also need to establish that the correlation holds when moving from laboratory videos to real-world production data.
But the main value of the MSU research is already clear. It gives developers more control. AI becomes less like magic and more like an instrument that can be understood and managed. That kind of foundation – based on understanding rather than blind replication – is what can underpin Russia’s technological sovereignty in AI.
