Systems and methods for training object detection models using adversarial examples

A computing system comprising:

• one or more processors;
• and
• one or more non-transitory computer-readable media that collectively store instructions that, when executed by the one or more processors, cause the computing system to perform operations, the operations comprising:
• obtaining data indicative of a training scene comprising a plurality of three-dimensional scene datapoints representing an environment;
• identifying a target object within the training scene, wherein the target object is represented by a plurality of three-dimensional target object datapoints;
• obtaining an adversarial object represented by a plurality of three-dimensional adversarial object datapoints;
• determining, by the computing system, an adversarial location for the adversarial object within the training scene based, at least in part, on the target object;
• generating data indicative of a modified training scene based, at least in part, on the training scene, the target object, the adversarial object, and the adversarial location, wherein the modified training scene comprises a three-dimensional representation of the adversarial object on the target object;
• determining a detection score by inputting the modified training scene to a machine-learned object detection model;
• and
• training the machine-learned object detection model based, at least in part, on the detection score.

A computing system, comprising:

• a training database comprising scene data and data indicative of an adversarial object, wherein the scene data comprises a plurality of training scenes, each training scene comprising a plurality of three-dimensional datapoints representing an environment;
• a machine-learned object detection model configured to receive a respective training scene of the plurality of training scenes, and in response to receipt of the respective training scene, output a respective detection output indicative of one or more training objects within the respective training scene;
• a memory that stores a set of instructions;
• and
• one or more processors which are configured to use the set of instructions to:
• obtain data indicative of a first training scene and the adversarial object from the training database;
• identify a first target object within the first training scene;
• determine a first adversarial location for the adversarial object within the first training scene based, at least in part, on the first target object;
• generate a first modified training scene based, at least in part, on the first training scene, the first training object, the adversarial object, and the first adversarial location, wherein the first modified training scene comprises a three-dimensional representation of the adversarial object on the first target object within the first training scene;
• determine a detection score by inputting the first modified training scene to the machine-learned object detection model;
• and
• generate an updated adversarial object by updating one or more of a plurality of adversarial object parameters of the adversarial object based, at least in part, on the detection score.