The analysis performed by a Reality Analysis job usually requires one or more Machine Learning models (e.g. a deep learning neural network). We call them ContextDetectors.
ContextDetectors are specific to Reality Analysis and provided by Bentley. They are classified in different categories, each Reality Analysis job type requiring a given category. Here are the categories used by current Reality Analysis jobs:
- photo object detector: ML model detecting objects as boxes in a photo.
- photo segmentation detector: ML model classifying pixels in a photo (semantic segmentation).
- orthophoto segmentation detector: ML model classifying pixels in an orthophoto (semantic segmentation). The difference with a photo segmentation detector is that pixel resolution is taken into account to improve detection.
- point cloud segmentation detector: ML model classifying points in a point cloud (semantic segmentation)
ContextDetectors are the basic blocks of Reality Analysis: limiting to these categories does not mean that analysis is limited to these basic tasks. For instance, 3D object detection is available but uses either a photo object detector or a point cloud detector underneath. Mesh analysis is also available, etc.
ContextDetectors available for download on the iTwin Capture detectors download page. To use one of them with Reality Analysis, just upload it to Reality Management Service in your Enterprise projects.