Questions & Answers

Please find below the list of frequent questions addressing common concepts, crucial details and some critical aspects of MetaVision's work.

What makes your approach special?

MetaVision's approach combines traditional but re-implemented from scratch computer vision algorithms (optimized for ultra-high-definition streams) and pre-trained machine learning to achieve better accuracy and significantly decrease the cost of data labelling.

Current implementation significantly speeds up the labelling process, producing better results for just a fraction of a cost.

Detailed segmentation example

The idea is to make the solution self-learning and mature enough to deal with complex scenes. Check the Roadmap

What tool do you use to build training sets?

MetaVision's toolset includes several in-house developed solutions, including the segmentation pipeline (based on a custom segmentation approach), the validation pipeline (used to make produced results consistent) and so-called meta-vision pipeline (designed to properly handle scene evolution over time).

Semantic segmentation

Do you use OpenCV?

No. OpenCV is unbelievably cool, but there is a need to have a bit more specialized and way faster implementation to deal with 4K or 8K video streams.

Detailed segmentation

Why do you only use video as an input?

Video stream segmentation along with detailed frame segmentation and time-based analysis can be used to make scene interpretation way more accurate and complete. Moreover, ultra-high-definition video streams transfer lots of useful information/insights.

Video stream segmentation

Video streams also provide so-called natural variance for the objects of interest, so it is easier to get better accuracy avoiding any kind of artificial data augmentation.

Do you plan to provide pre-trained models?

Yes. Pre-trained models will be provided as a natural continuation of MetaVision's development strategy.

Video stream segmentation

It is worth mentioning all available training sets are basically low-level "building blocks", providing an option of building a highly customized final set for training a pretty specific model.

Did you compare your results with available datasets?

Yes. Here is a very brief comparison against Open Images Dataset V6.

Brief comparison against Open Images Dataset V6

In-depth comparison, benchmarks and tests will be provided.

Do you plan to opensource your solution?

Long story short, yes. There is a plan to release a simplified segmentation pipeline as a standalone version.

Are you going to update your training sets periodically?

Yes, training sets are periodically updated and extended.

Updated training-set

©   MetaVision — 4K/8K video stream segmentation