MTS AI explains on Habr how it is changing the training recipe for Luna Line segmentation models

The MTS AI team has published the second part of its Habr series on finding the best training recipe for CV models in the Luna Line family. This time, the focus is on segmentation models.

What the first part on classification showed

In the first part of the series, the MTS AI team covered training classification models in Luna Line and arrived at a unified pipeline that works across different tasks of this type. Once that work was done, the authors were left dissatisfied with their methodology. They weren't sure the chosen model family was actually performing at its best, rather than simply being squeezed into a one-size-fits-all pipeline designed for convenience and versatility.

That question — whether the family was really reaching its full potential or being artificially held back by a shared framework — became the starting point for the second part of the series, which shifts from classification to segmentation.

How the approach changed for segmentation

• Part 1 of the series covers training classification models in Luna Line and is available on the MTS AI company blog on Habr
• Part 2 covers segmentation, a different type of computer vision task: instead of assigning a single label to the whole image, the model outlines specific objects at the pixel level
• For segmentation, the team chose a different approach than it used for classification, to avoid repeating the earlier constraint of a single universal pipeline
• Luna Line is a broader MTS AI initiative to find a universal recipe for training CV models across different task types

The authors are explicit about why they changed course. The goal isn't just to lock in a working pipeline for the sake of reporting, but to make sure the chosen architecture actually reaches its real potential on a specific task, instead of conforming to constraints inherited from a different stage of the work.

Why classification and segmentation can't be trained the same way

Classification answers the question "what does the image show overall," while segmentation has to precisely trace the boundaries of each individual object at the pixel level. These tasks differ fundamentally in complexity and structure: segmentation places higher demands on feature resolution, pixel-level class balancing, and the design of the network's head. A unified pipeline tuned for classification risks ignoring these differences and dragging down the final quality of segmentation models.

Why the Luna Line series exists at all

The idea behind Luna Line is to find not a single specific model, but a reproducible method for selecting and training CV models that can be applied to new tasks without lengthy manual searches over architectures and hyperparameters. For the MTS AI team, this is a practical engineering problem: a large company runs dozens of CV scenarios at once, and searching for the optimal configuration from scratch every time eats up a lot of specialists' time.

That's why the interim conclusion from the classification part — doubt that a unified pipeline fully unlocks the model family's potential — mattered. It changed how the team tackled the next task in line, rather than simply copying the classification methodology over to segmentation without checking it.

What it means

Segmentation and classification are fundamentally different computer vision tasks, and trying to train both through the same universal process risks lowering the quality of each. MTS AI's decision to rethink its approach specifically for segmentation shows that a "universal recipe" in CV engineering isn't a single fixed pipeline. It's closer to a methodology that has to be deliberately adapted to the type of task at hand.