MTS AI Discussed on Habr How Training Methods for Luna Line CV Segmentation Models are Changing

The MTS AI team published the second part of an analysis on Habr about finding the optimal training recipe for CV models in the Luna Line lineup — this time discussing segmentation models.

What the first part on classification showed

In the first part of the series, the MTS AI team analyzed training of classification models in Luna Line and derived a single pipeline applicable to different tasks of this type. After completing this work, the authors were left with a sense of dissatisfaction with the chosen methodology: they were not sure that the selected family of models really showed its maximum potential, rather than being constrained within the framework of a unified pipeline invented for convenience and universality.

This question — whether the family truly reveals its full potential or is artificially limited by common frameworks — became the starting point for the second part of the series, devoted not to classification, but to segmentation.

How the approach changed for segmentation

• Part 1 of the series is devoted to training classification models in Luna Line and is available on the MTS AI company blog on Habr
• Part 2 is devoted to segmentation — a different type of computer vision task where the model does not label the entire image with a single label, but instead traces the contours of specific objects at the pixel level
• For segmentation, the team decided to act differently than in the case of classification, to avoid repeating the previous limitation of a single universal pipeline
• The Luna Line project is a joint initiative of MTS AI to find a universal training recipe for CV models for different types of tasks

The authors directly formulate the motivation for changing their approach: the goal is not simply to fix a working pipeline for the sake of reporting, but to ensure that the selected architecture truly reveals its real potential on the specific task, rather than adapting to someone else's frameworks inherited from another stage of work.

Why you cannot train classification and segmentation the same way

Classification answers the question "what is shown in the picture as a whole," while segmentation must precisely outline the boundaries of each individual object at the pixel level. These are fundamentally different tasks in terms of complexity and structure: segmentation has higher requirements for feature resolution, class balancing at the pixel level, and network head architecture. A unified pipeline optimal for classification risks not taking these features into account and lowering the ultimate quality of segmentation models.

Why the Luna Line series appeared in the first place

The idea of Luna Line is to find not a single specific model, but a reproducible method of selecting and training CV models that can be applied to new tasks without lengthy manual iteration of architectures and hyperparameters. For a team from MTS AI, this is a practical engineering problem: a large company has dozens of CV scenarios in progress simultaneously, and searching for the optimal configuration from scratch each time is expensive in terms of specialist time.

That is why the intermediate conclusion after the classification part — doubt that a unified pipeline fully reveals the family's potential — turned out to be important: it changed how the team approached the next task in the queue, instead of simply copying the classification methodology to segmentation without verification.

What this means

Segmentation and classification are different in nature as computer vision tasks, and attempting to train them through the same universal process risks lowering the quality of each. MTS AI's decision to reconsider the approach specifically for segmentation shows that a "universal recipe" in CV engineering is not a single fixed pipeline, but rather a methodology that must be consciously adapted to the type of problem being solved.