Understanding performance of any machine learning solution is of fundamental importance. Therefore, perClass Mira 1.1 introduces confusion matrix visualization visualizing error brake-down within individual classes. Both normalized and absolute confusion matrices are available together with convenient additional information such as per-class errors or precision (ratio of a correct acceptance and all deciions of a given class).

Because static performance visualization does not allow us to improve, we have developed a fully interactive confusion matrix for perClass Mira. User may lower errors at specific entries (or for the entire class) using sliders. In addition, multiple constraints may be easily created by double-clicking the entries. Then only solutions fulfilling all constraints are considered. Constraints may be adjusted live (via Ctrl+mouse wheel) and enabled/disabled at will. This provides a strong yet easy-to-use environment for performance fine-tuning. Any changes in the confusion matrix are directly reflected in the model decisions. In this way, we may quickly create solutions fulfilling specific application requirements.

Additional new functionality in the 1.1 release:

• An RGB preview constructed from user-adjustable R, G and B band lines in the spectral plot. This significantly simplifies labeling of complex scenes where single-band view does not offer sufficient understanding of class localization

• undo/redo functionality for label painting
• cropping of large cubes. It is possible to crop a cube and save the project. Once opened again, the full cube will be loaded and cropped automatically - there is no need to save the cropped image data separately from existing data storage. It is also possible to add the original cube and crop another region or use the entire scan for testing.
• specific classes may be excluded from training
• more in release notes

Want to try out this new functionality? Request a demo!