Contains samples for implementing Windows Skills by extending the preview base API and using exsting skill packages
MIT
This code repo is now deprecated and the project is unsupported
AI Skills for Windows
Summary
Implementing and integrating efficient AI and Computer Vision (CV) solutions is a hard task for developers. The industry is moving at a fast pace and the amount of custom-tailored solutions coming out makes it almost impossible for app developers to keep up easily. We preview a base framework to expose AI solutions called AI Skills for Windows as well as pre-built common solutions (i.e. detection, classification, segmentation, etc.) developed by Microsoft and partners.
The AI Skills for Windows framework is meant to standardize the way AI and CV is put to use within a Windows application (i.e.: UWP, Desktop Win32, .Net Core 3.0+) running on the edge. It aims to abstract away the complexity of AI techniques by simply defining the concept of skills which are modular pieces of code that process input(s) and produce output(s). The implementation that contains the complex details (i.e. pre and post processing of data, model inference, algorithm, transcoding, applying the right heuristics, etc.) is encapsulated by WinRT APIs that inherits the base class present in the Microsoft.AI.Skills.SkillInterface namespace, which leverages built-in Windows primitives which in-turn eases interop with built-in hardware acceleration leveraged by frameworks such as Windows ML. All AI Skills for Windows derivatives follow the same programatic paradigm and flow from a developer consumer standpoint: if you understand how to use one AI Skill for Windows, you understand how to use them all. (See key AI Skills API concepts)
While this release focuses on vision-oriented scenarios and primitives, this API is meant to accommodate any kind of input and output variable and a wide range of scenarios (Vision, Audio, Text, etc.). Any developer can extend this API set and expose their own AI skills. See skills released by Intel
If you are looking for the earlier preview release samples and documentation, we archived them in a branch here: Preview branch
Searches an image for quadrilateral shapes and returns the coordinates of their corners if found. The LiveQuadDetector is a stateful version of the QuadDetector that attempts to detect only 1 quadrangle and keeps track of the previous quad detected to be used as guide which optimizes tracking performance as new frames are bound over time. This is well suited for most scenarios operating over a stream of frames over time. QuadDetector can be set to detect more than 1 quadrangle and will search the whole frame everytime unless a previous quadrangle is provided.