Neurotechnology announced SentiSight 3.0, a Software Development Kit (SDK) for universal object recognition.
The object recognition algorithms now enable an even broader range of recognition capabilities for applications as varied as manufacturing, artificial intelligence, searches for identifiable marks and even place recognition.
SentiSight provides enhanced 2D and 3D object recognition quality using still or video images from most digital cameras, including Webcams. It can detect and recognize whether a particular rigid object, such as a product, logo or building, is in a scene and identify its specific location in that scene.
It can also count the number of specific identified objects in a scene and can compare two photographic images to provide place recognition, based on objects within the picture.
The new shape-based algorithm is suitable for localization and recognition with objects that have distinguishable external or internal edges. The algorithm is fully tolerant of in-plane rotation, up to 15-20 degrees of out-of-plane rotation (such as from frontal to profile) and a wide range of changes in scale.
It can handle occlusion of up to 50% as long as enough unique edges of the object are still visible. Multiple views can be added to the object model to provide even more reliable recognition or better out-of-plane rotation tolerance. The shape-based algorithm offers enhanced recognition at near real-time performance in many conditions.
The enhanced local-feature-based algorithm offers even faster recognition for objects that have clear and stable local features, such as bank notes, brand labels on packaging, logos, etc. In addition to 30-40 times faster recognition speeds (when using a quad-core processor), the “learning” mode, where objects or images are presented to the system, now takes up to 40% less time and the model size is two times smaller.
The overall quality of recognition is improved over the previous version, with a 10-30% reduction in the false rejection rate.
The new tracking algorithm provides enhanced tracking of objects in front of complex backgrounds and performs automatic, reliable tracking of fast-moving objects after they have been recognized. The algorithm can track local-feature-rich as well as edge-feature-rich objects.