VisionLabs finds a way to train CV models on 50 images without coding

VisionLabs (part of MWS AI) has published the first part of a technical series on the development of Luna Line. Anastasia Belozerova, team lead of the research team, describes in detail how the engineers searched for a "universal recipe" for training classification models and what came of it.

A platform for non-specialists

Luna Line is a no-code computer vision platform built by VisionLabs for users with no programming skills. An agronomist, a production engineer, a quality inspector or a logistics specialist can label a set of images, click a button and get a ready-trained CV model for their specific task.

The key product requirement is that it must work with very small datasets. The team targets a scenario starting from 50 images, which is a real constraint for most industrial enterprises, which have neither the time nor the specialists to label thousands of examples. This constraint drives the whole difficulty of the engineering problem: standard recipes from academic datasets do not work here.

The search for a universal recipe

The team set itself an ambitious goal: to find a single training configuration that reliably delivers good quality on any classification dataset. If such a recipe exists, the platform can work as a "black box": the user brings the data, and the system configures everything else on its own.

To test hypotheses, the team built a strict experimental methodology:

• Hypothesis: a specific configuration is formulated: backbone architecture, augmentations, learning rate schedule, optimizer
• Testing: the configuration is run on several datasets of different character, size and domain
• Comparison: results are compared with baselines and earlier hypotheses
• Decision: the configuration is accepted as "universal" or rejected based on the analysis

This approach lets the team move systematically rather than by random trial and error, gradually narrowing the search space to solutions that actually work.

Why the methodology changed

One of the series' key findings was unexpected: configurations that proved best for classification did not transfer automatically to segmentation. When the task class changed, the team found that some previously accepted decisions had to be revisited from scratch.

"I will explain what hypotheses we put forward, how we tested them and why we revised the experimental methodology when moving from one task to another,"

Belozerova says in her announcement.

This led to the conclusion that a universal recipe exists only within a single class of tasks. There is no direct transfer between tasks, but there is something valuable: the search methodology itself proved transferable. It can be applied to segmentation and detection, each time starting the experimental "branch" anew, but with a properly built process for verifying hypotheses.

The second part of the series, on segmentation, will be published later.

What this means

VisionLabs shows an engineering openness that is rare on the Russian ML market: the company publishes not marketing talking points but honest chronicles of experiments, including negative results and revised hypotheses. For specialists building their own MLOps pipelines or no-code tools, this is a valuable practical reference, especially for work with small datasets, where academic benchmarks are almost useless.