Training neural networks for vehicle re-identification

A method comprising:

• computing, using a plurality of instances of a neural network, data representative of embeddings corresponding to an embedding space, the embeddings including a first embedding corresponding to an anchor image depicting a vehicle and a plurality of second embeddings corresponding to images depicting the vehicle;
• determining, based at least in part on the data, a distribution corresponding to distances between the first embedding and one or more second embeddings of the plurality of second embeddings;
• based at least in part on the distribution, selecting a second embedding from the plurality of second embeddings associated with a highest probability from the distribution;
• and
• training the neural network using the second embedding.

A system comprising:

• one or more cameras to generate data representative of an environment;
• one or more processing units including processing circuitry to:
• compute, using a neural network, data representative of embeddings corresponding to an embedding space, the embeddings including a first embedding corresponding to an anchor image depicting a vehicle and a plurality of second embeddings corresponding to images depicting the vehicle;
• and
• train the neural network based at least in part on a second embedding selected from a distribution corresponding to distances between the first embedding and one or more second embeddings of the plurality of second embeddings and based at least in part on probabilities associated with the distribution.