Abstract: Provided are methods for automated verification of annotated sensor data, which can include receiving annotated image data associated with an image, wherein the annotated image data comprises an annotation associated with an object within the image, determining an error with the annotation based at least in part on a comparison of the annotation with annotation criteria data associated with criteria for at least one annotation, determining a priority level of the error, and routing the annotation to a destination based at least in part on the priority level of the error. Systems and computer program products are also provided.

Abstract: A system receives a 3D image having multiple data points, and uses one or more filters, such as a distance filter, map filter, and/or height filter to remove certain 3D data points from the image. The system may group the data points and annotate them to identify unknown or unclassified objects within the image.

Abstract: Provided are methods for offline perception motion inference, which can include obtaining map data indicative of an environment and obtaining data associated with at least one agent. The method can include determining a trajectory for the agent and matching the trajectory of the agent with a lane connector. The method can also include determining a traffic light parameter. Systems and computer program products are also provided.

Abstract: Among other things, techniques are described for cross-modality active learning for object detection. In an example, a first set of predicted bounding boxes and a second set of predicted bounding boxes is generated. The first set of predicted bounding boxes and the second set of predicted bounding boxes are projected into a same representation. The projections are filtered, wherein predicted bounding boxes satisfying a maximum confidence score are selected for inconsistency calculations. Inconsistencies are calculated across the projected bounding boxes based on filtering the projections. An informative scene is extracted based on the calculated inconsistencies. A first object detection neural network or a second object detection neural network is trained using the informative scenes.

Abstract: Provided are methods for customized tags for annotating sensor data, which can include receiving sensor data captured during a plurality of sensor data capture sessions, processing the sensor data using a plurality of machine learning models to identify a plurality of capture session collections represented in the sensor data, filtering the sensor data based at least partly on a user-specified category of the plurality of categories of capture session to identify a capture session collection, of the plurality of capture session collections, representing sensor data of one or more sensor data capture sessions that conforms to the user-specified category, and transmitting the sensor data of one or more sensor data capture sessions that conforms to the user-specified category to an end user computing device. Systems and computer program products are also provided.

Abstract: Provided are methods for automated verification of annotated sensor data, which can include receiving annotated image data associated with an image, wherein the annotated image data comprises an annotation associated with an object within the image, determining an error with the annotation based at least in part on a comparison of the annotation with annotation criteria data associated with criteria for at least one annotation, determining a priority level of the error, and routing the annotation to a destination based at least in part on the priority level of the error.

Abstract: Among other things, techniques are described for cross-modality active learning for object detection. In an example, a first set of predicted bounding boxes and a second set of predicted bounding boxes is generated. The first set of predicted bounding boxes and the second set of predicted bounding boxes are projected into a same representation. The projections are filtered, wherein predicted bounding boxes satisfying a maximum confidence score are selected for inconsistency calculations. Inconsistencies are calculated across the projected bounding boxes based on filtering the projections. An informative scene is extracted based on the calculated inconsistencies. A first object detection neural network or a second object detection neural network is trained using the informative scenes.

Abstract: Among other things, we describe techniques for automatically selecting data samples for annotation. The techniques use bounding box prediction based on a bounding box score distribution, spatial probability density determined from bounding box sizes and positions and an ensemble score variance determined from outputs of multiple machine learning models to select data samples for annotation. In an embodiment, temporal inconsistency cues are used to select data samples for annotation. In an embodiment, digital map constraints or other map-based data are used to exclude data samples from annotation. In an exemplary application, the annotated data samples are used to train a machine learning model that outputs perception data for an autonomous vehicle application.

Abstract: Provided are methods for customized tags for annotating sensor data, which can include receiving sensor data captured during a plurality of sensor data capture sessions, processing the sensor data using a plurality of machine learning models to identify a plurality of capture session collections represented in the sensor data, filtering the sensor data based at least partly on a user-specified category of the plurality of categories of capture session to identify a capture session collection, of the plurality of capture session collections, representing sensor data of one or more sensor data capture sessions that conforms to the user-specified category, and transmitting the sensor data of one or more sensor data capture sessions that conforms to the user-specified category to an end user computing device. Systems and computer program products are also provided.

Abstract: Among other things, we describe techniques for automatically selecting data samples for annotation. The techniques use bounding box prediction based on a bounding box score distribution, spatial probability density determined from bounding box sizes and positions and an ensemble score variance determined from outputs of multiple machine learning models to select data samples for annotation. In an embodiment, temporal inconsistency cues are used to select data samples for annotation. In an embodiment, digital map constraints or other map-based data are used to exclude data samples from annotation. In an exemplary application, the annotated data samples are used to train a machine learning model that outputs perception data for an autonomous vehicle application.