Data Sheets

MXR-DS-deepcore 10/20 Tagging and validation. DeepCore provides a quick and efficient tagging and validation environment to build ML models. This enables customers to evaluate how a model performs against ground truth data for a given set of images and identify model performance issues. Training data and model catalogs. DeepCore training data and model catalogs allow users to search, filter, compare, upload, export and geospatially and temporally visualize training data and models respectively. The training data catalog is mapped to the xView data schema and ontology, a database of over 3 million training samples created by Maxar. The catalog not only provides a place for users to view and compare models but will continue to be updated through a model run feedback loop. Users will be able to compare model selection, model performance and other data analysis metrics. DeepCore Vision analytical UI. The DeepCore Vision UI visualizes object detections within a user's programs and projects plus any publicly shared detections. By showing them geospatially and temporally, users can validate model outputs, conduct basic data analysis and investigate more details about the object detected. This allows users to see what models would have detected at lower or higher precision and recall. Vision addresses broad area search, rare object and trend analysis use cases. Vision is written in React.js, a modular JavaScript framework completely compatible with today's modern browsers. Our back-end services use PostgreSQL databases with the enabled PostGIS extension for easy geospatial calculations and support. The microservices were developed on various web technologies, including Java Spring Boot, Node.js and Python. Easy compatibility without hardware requirements. DeepCore has supported over 100 accessible algorithms available from such sources as Caffe, Caffe2 via ONNX, TensorFlow and PyTorch via Torch Script. Since the web components require no hardware, DeepCore uses standard hardware and networks that can be cloud-based, on-premise or a hybrid of the two. Diverse imagery sources. DeepCore imagery services have connected to select AOIs for iSpy, OMAR, NCL and ODS to provide clients scalable ML processing for imagery (airborne, satellite, hand-held) and signals. DeepCore performs inferencing that allows collection managers to specify a location on Earth, objects to find and start/ end dates for processing to support the CV model development process. The U.S. government retains SBIR data rights for DeepCore Suite, making it scalable and free of cost. info@maxar.com maxar.com What does DeepCore provide? FEATURES continued ■ Ability to visualize millions of object detections in the DeepCore Vision UI, or integrate with ELT and GIS web and desktop soware □ Labeled training data in 130 object types using formatted ontology for easy ETL □ Integration with Open Source Soware Image Map orthorectification processes for high geospatial accuracy □ Bidirectional translation between unorthorectified and orthorectified detections and visualization ■ Support for satellite, airborne, drone and terrestrial sources □ Electro-optical, synthetic aperture radar and multispectral imagery □ Some full-motion video DEEPCORE SUITE COMPONENTS ■ Tagging and validation ■ Training data catalog ■ Model catalog ■ Collection manager ■ Inference server ■ Imagery services REQUIREMENTS USED BY DEEPCORE Hardware: Cloud-based, on-premise or combination Nvidia GPUs (DeepCore Server) Operating systems: Linux Soware: Linux, Java 8 and a compatible C++ runtime (DeepCore Server) Back-end services: PostgreSQL databases with enabled PostGIS extension DeepCore vision UI. Training Object Submit Content management/ AIIRS Vision Validate and Verify (Verifi ed results are added to training data to generate improved machine learning model)

• Cover