Automated annotation techniques

A method of suggesting an annotation for an object in a sensor image, comprising:

• generating a set of predictions for the sensor image by applying an annotation model to the sensor image, each prediction indicating a possible location of the object on the sensor image,
• displaying the sensor image on a client device associated with a human operator, wherein the set of predictions are not displayed on the sensor image;
• responsive to the human operator interacting with the sensor image on the client device, identifying the suggested annotation from the set of predictions that corresponds to a location of the user interaction on the sensor image;
• and
• displaying the suggested annotation on the client device.

A method of updating an annotation for an overlapping set of objects in a sensor image, comprising:

• displaying the sensor image on a client device associated with a human operator, wherein the annotation for the sensor image groups the overlapping set of objects into a single label;
• receiving input from the client device for one or more annotations that each identify a corresponding object in the overlapping set of objects;
• responsive to receiving the input from the client device, determining an updated set of annotations for remaining objects that have not been annotated by the human operator, wherein the updated set of annotations are determined based on the original annotation for the sensor image and the one or more annotations identified by the human operator;
• and
• displaying the updated set of annotations on the client device.

A system comprising one or more processors and computer storage media storing instructions that when executed by the one or more processors, cause the processors to perform operations comprising:

• obtaining a sequence of sensor images from one or more sensors, the sequence of sensor images including sensor measurements that sense a surrounding environment of the one or more sensors;
• generating a set of annotations that indicate locations of objects in the sequence of sensor images, the set of annotations generated by applying a bidirectional machine-learned model to the sequence of sensor images, comprising:
• generating a set of forward estimations for the sequence of sensor images by propagating a first sensor image of the sequence through the bidirectional machine-learned model in a forward direction,
• generating a set of backward estimations for the sequence of sensor images by propagating a last sensor image of the sequence through the bidirectional machine-learned model in a backward direction, and
• combining the set of forward estimations and the set of backward estimations to generate the set of annotations;
• and
• training a computer model using the annotated sequence of sensor images.