Labeling

Use the Multi-Sensor Labeler app to interactively label ground truth data across multiple time-synchronized signals representing the same scene. You can label videos, image sequences, and lidar point cloud signals simultaneously. The app supports a wide range of ROI label types, including rectangles, rotated rectangles, cuboids, projected cuboids, lines, polygons, pixel labels, and semantic point labels. You can also apply scene labels to classify entire frames across all signals. To get started, see Get Started with Multi-Sensor Labeler.

To accelerate the labeling process, the app provides built-in automation algorithms for object tracking and temporal interpolation across frames. You can also develop and import custom automation algorithms. For more information, see Automate Labeling for Multi-Sensor Data.

Once labeling is complete, export the labeled data as a groundTruthMultiSensor object and create training data for deep learning workflows. Use pointCloudObjectDetectorTrainingData to create training data for 3-D object detection networks such as PointPillars, or use semanticPointLabelTrainingData to create paired datastores for training 3-D semantic segmentation networks such as RandLA-Net. For more information, see Export Multi-Sensor Ground Truth and Create Training Data.

To define label definitions programmatically outside the app, use the labelDefinitionCreatorMultiSensor object to create, inspect, and manage label hierarchies including sublabels and attributes. You can build groundTruthMultiSensor objects programmatically by combining data sources with label definitions, ROILabelData objects for ROI labels, and SceneLabelData objects for scene labels. You can also import programmatically created ground truth objects into the app for further review and labeling.