Data alone is almost worthless, but orderly and flawlessly labeled data is an essential value for modern companies.
You can create value with us
Dynamic user interface
Our annotator can be perfectly customized with the help of dynamically editable screen panels, so you can always work with tools tailored to the problem. These changes can be saved so later you can retrieve the views that best fit the task with one click.
The most basic task is to label existing data based on specific quality characteristics. Often it is enough to classify the examined samples into OK/NOK classes, but it is possible to form complex hierarchical classes too. In solving the classification tasks, we strived to provide all users with the most convenient and fastest possible method.
Using the “Quick Tag” tool, we can customize the user interface for the classification. The large buttons that can be placed in any position provide convenient use, especially when used with a touch screen.
If a touch screen interface is not available, you don’t have to deal with mouse positioning. You can assign the annotation labels to keyboard shortcuts, so you can do the task with even one hand.
Segmentation - masking techniques
In many cases, it is not enough to provide qualitative or quantitative annotations, the accurate segmentation can be an essential element of image recognition too. Most AI models that mimic vision require pixelwise annotation. This manual task can be responsible for 90% of the cost and time required for AI projects. At Machine Intelligence, we know that significant time can be saved on this task, so we paid special attention to developing the most efficient segmentation tools.
With the classic tool, you can select exactly what you want. In proportion to the number of control points, we can achieve ever finer resolution and precision. You can use a line consisting of straight sections or interpolated curves. With the help of the accessories, the already drawn polygons can be recycled and transformed if required.
If the masking task is complex, but you don’t want to paint every single pixel by hand, the superpixel masking can be a great help, as you can paint statistically similar groups of pixels as a single pixel. The sensitivity/resolution of the superpixel method can be changed dynamically on the fly.
In the special procedure, we only need to give the algorithm a few examples of which areas we want to draw with which mask. The rest is done automatically by finding the statistically most probable boundaries. Providing a sample can be not just a one-time operation, but can be supplemented if we are not yet satisfied with the intermediate result.
Need to locate objects or errors? With the help of our annotator, you can mark any point-like or rectangular areas, which can be modified, moved, selected or even transformed afterwards.
The special point tag annotation of MInD allows very accurate localization. In contrast to the object detection-based versions that work with anchors and local regressions, the segmentation-based localization network provides an order of magnitude more accurate solution.
Box tagging can already be called a classic, as it has been used as an input for object detection for nearly half a decade.