Human vision works well interpreting complex and quite often unstructured scenes. Machine vision solutions are far from that maturity level but can deal with less complex scenes faster, providing accurate and repeatable results which enables automation of many production scenarios. Moreover, properly implemented machine vision systems can inspect small details which are practically inaccessible to the human eye.


Machine vision is the approach and technology intended to achieve accurate scene segmentation/decomposition to make automatic inspection and control possible. A common result provided by automatic inspection solution is pass/fail decision, which might be limiting in case of a complex problem. To address that MetaVision solution provides additional details including accurate geometry, semantic relations, behavioural details, etc.

Additionally, numerical measurements, detected orientation and classification results can be provided. Exact position and orientation, for example, are usually used to remove misaligned components preventing a potentially critical problem.

Detailed segmentation of a flower

Video streams segmentation

High-definition video streams transfer lots of useful information/insights. Extraction of valuable details from the stream can be viewed as an important area of application in machine vision. The initial step and crucial basis in video stream analysis is frame segmentation.

Detailed segmentation of a flower

Image segmentation has been a proper way to solve many business problems. Accurate and robust video stream segmentation provides a straightforward way to automate a wide variety of processes. Complex shape recognition, including tricky internal details, is crucial for many applications.

Product inspection problem

Let's use the product inspection problem as a good illustration helping to understand the way a machine vision solution works. Using a proper quality input the solution can provide detailed visual inspection using any type of pre-trained o predefined model. Fixed camera placement, as well as predefined scenes parameters, allows providing accurate defects detection and measurement (based on pre-defined quality gates).

Any production line with a repetitive process might benefit from introducing a machine vision solution. Such solutions are being widely used in medical, automotive, food and other areas. Compared to human inspection machine vision is faster, provides results continuously without any kind of quality degradation.

Cookie detailed segmentation

Video stream segmentation along with detailed frame segmentation and time-based analysis can be used to make any business way more efficient, transparent and smarter. Automatic inventory analysis, efficient diagnostic, defect recognition and measurement are just a few examples.

4K/8K stream processing

MetaVision's approach combines traditional re-implemented from scratch computer vision algorithms to deal with 4K and 8K streams and pre-trained machine learning to achieve impressive accuracy.

Cookie detailed segmentation

Segmentation pipelines

Several general-purpose segmentation pipelines have been developed, due to the fact, there is no general solution, providing accurate and robust results. Segmentation techniques provide a simple way to extract so many valuable details.

Segmentation is about finding exact objects boundaries, internal details and separating them from each other, better overall quality and higher resolution might significantly improve segmentation results.

This is the main reason MetaVision's solutions are implemented in a way 4K or 8K pixel-level precision is not a problem.

Complex flower segmentation

The approach taken is not only about extracting accurate geometry of the object, but also about building a modular and extensible machine vision system, capable of dealing with multiple input video streams providing semantic segmentation and extracting all the relevant details for any given particular scene.

Semantic segmentation

Semantic segmentation based on convolutional neural networks under a complex background might be tricky also providing inaccurate results. To improve the accuracy and maximize confidence a hybrid approach is proposed combining both flexibilities of machine learning and the accuracy of traditional algorithms.

Semantic segmentation

Tradition neural networks used to have various segmentation issues including rough edges and poor segmentation accuracy. Hybrid approach implemented by MetaVision provides impressive accuracy and literally "vectorizes" input stream.

Semantic segmentation

©   MetaVision — 4K/8K video stream segmentation