US 11,842,282 B2
Neural networks for coarse- and fine-object classifications
Junhua Mao, Palo Alto, CA (US); Congcong Li, Cupertino, CA (US); and Yang Song, San Jose, CA (US)
Assigned to Waymo LLC, Mountain View, CA (US)
Filed by Waymo LLC, Mountain View, CA (US)
Filed on Jun. 9, 2022, as Appl. No. 17/836,287.
Application 17/836,287 is a continuation of application No. 17/118,989, filed on Dec. 11, 2020, granted, now 11,361,187.
Application 17/118,989 is a continuation of application No. 16/229,332, filed on Dec. 21, 2018, granted, now 10,867,210, issued on Dec. 15, 2020.
Prior Publication US 2022/0374650 A1, Nov. 24, 2022
Int. Cl. G06V 20/56 (2022.01); G06N 3/084 (2023.01); G05D 1/02 (2020.01); G06N 3/08 (2023.01); G06F 18/20 (2023.01); G06F 18/25 (2023.01); G06F 18/21 (2023.01); G06F 18/214 (2023.01); G06F 18/2431 (2023.01); G06N 3/044 (2023.01); G06V 30/19 (2022.01); G06V 30/24 (2022.01); G06V 10/764 (2022.01); G06V 10/80 (2022.01); G06V 10/82 (2022.01); G06V 20/58 (2022.01)
CPC G06N 3/084 (2013.01) [G05D 1/0221 (2013.01); G05D 1/0231 (2013.01); G06F 18/217 (2023.01); G06F 18/2148 (2023.01); G06F 18/2431 (2023.01); G06F 18/25 (2023.01); G06F 18/285 (2023.01); G06N 3/044 (2023.01); G06N 3/08 (2013.01); G06V 10/764 (2022.01); G06V 10/809 (2022.01); G06V 10/82 (2022.01); G06V 20/56 (2022.01); G06V 20/58 (2022.01); G06V 30/19173 (2022.01); G06V 30/2504 (2022.01)] 20 Claims
OG exemplary drawing
 
1. A method for training an object classifier neural network, comprising:
training, in a first phase, the object classifier neural network to generate course-object classifications with a first set of training data, the first set of training data including a first plurality of training examples, each training example in the first set of training data being labeled with a coarse-object classification; and
training, in a second phase after completion of the first phase, the object classifier neural network to generate fine-object classifications with a second set of training data, the second set of training data including a second plurality of training examples, each training example in the second set of training data being labeled with a fine-object classification,
wherein training the object classifier neural network in the second phase comprises, for each training example, adjusting parameters of one or more first encoding portions of the neural network without adjusting parameters of one or more second encoding portions of the neural network,
wherein the one or more first encoding portions of the neural network produce outputs for input to (i) a coarse-object classification portion of the neural network and (ii) a particular fine-object classification portion of the neural network assigned to the fine-object classification indicated by the label of the training example,
wherein the one or more second encoding portions of the neural network produce outputs for input to the coarse-object classification portion of the neural network but not the particular fine-object classification portion of the neural network.