Pre-labeling data with cuboid annotations

A computer-implemented method for labeling data, the method comprising:

• training a first machine learning (ML) model using a loss function evaluated by performing steps including:
• predicting, using the first ML model, a first cuboid annotation;
• determining a first distance between corners of a cuboid annotation included in training data and a second cuboid annotation that includes a size of the first cuboid annotation, a position of the cuboid annotation included in the training data, and a yaw of the cuboid annotation included in the training data;
• determining a second distance between corners of the cuboid annotation included in the training data and a third cuboid annotation that includes a position of the first cuboid annotation, a size of the cuboid annotation included in the training data, and a yaw of the cuboid annotation included in the training data;
• determining a third distance between corners of the cuboid annotation included in the training data and a fourth cuboid annotation that includes a yaw of the first cuboid annotation, a size of the cuboid annotation included in the training data, and a position of the cuboid annotation included in the training data;
• and
• determining an average of the first distance, the second distance, and the third distance;
• processing point cloud data using the first ML model and a second ML model to generate cuboid annotations of objects in the point cloud data by performing steps including:
• processing the point cloud data using the first ML model to determine cuboid tracks associated with the objects;
• projecting regions of the point cloud data including the objects to a plurality of two-dimensional (2D) views;
• and
• processing the plurality of 2D views and the cuboid tracks using the second ML model to determine updated cuboid tracks associated with the objects;
• determining a confidence threshold based on a count of the cuboid annotations;
• determining, based on the confidence threshold, one or more of the cuboid annotations to be displayed;
• and
• causing to be displayed the one or more of the cuboid annotations based on the determining, and at least one user interface element that permits a user to select, confirm, or modify the one or more of the cuboid annotations that is displayed.

A computer-implemented method for training a machine learning (ML) model, the method comprising:

• receiving training data including point cloud data and a labeled cuboid bounding at least one object in the point cloud data;
• and
• training a ML model using a loss function that is evaluated by performing steps including:
• predicting, using the ML model, a cuboid annotation, each of the cuboid annotation and the labeled cuboid indicating features of the bounding of the at least one object, the features including a size, a position, and a yaw,
• for each feature,
• determining an additional cuboid annotation based on the feature of the cuboid annotation and remaining features of the labeled cuboid, and
• determining a distance between the labeled cuboid and the additional cuboid annotation, and
• determining an average of the distances between the labeled cuboid and the additional cuboid annotation determined for each feature.