Object detection with cross-domain mixing
Implementations are described herein for improving unsupervised domain adaptation (UDA) by using improved adaptive teacher for object detection with cross-domain mix-up. In various implementations, cross-domain training of an object detection machine learning model may include: performing weak augmentation on images from a target domain D T to generate a first set of weakly augmented target domain images; perform strong augmentation on images from the source domain D S and images from the target domain D T to generate a second set of strongly augmented images; processing the second set of strongly augmented images to generate a third set of inter-domain mixes of the images from D S and D T ; and jointly train the object detection machine learning model, as a student machine learning model, with a teacher machine learning model using the first and third sets.
1 . A method for performing cross-domain training of an object detection machine learning model, the method implemented using one or more processors, comprising:
performing weak augmentation on images from a target domain (DT) to generate a first set of weakly augmented target domain images;
performing strong augmentation on images from a source domain (DS) and the images from the DT to generate a second set of strongly augmented images;
processing the second set of strongly augmented images to generate a third set of inter-domain mixes of the images from the DS and the DT; and
jointly training the object detection machine learning model, as a student machine learning model, with a teacher machine learning model using the first and third sets.
2 . The method of claim 1 , further including processing images from the second set to generate a fourth set of intra-domain mixes of the images from the DS.
3 . The method of claim 2 , wherein the object detection machine learning model is further trained using the fourth set.
4 . The method of claim 1 , wherein the jointly training includes updating weights of the teacher machine learning model based on an exponential moving average of weights of the object detection machine learning model.
5 . The method of claim 1 , wherein output generated using the teacher machine learning model includes pseudo labels that are used to train the object detection machine learning model.
6 . The method of claim 1 , further including performing adversarial training of the object detection machine learning model and a discriminator machine learning model, wherein the performing adversarial training includes applying the discriminator machine learning model to a feature embedding generated by the object detection machine learning model to generate a predicted classification of the feature embedding as being in the source DS or the DT.
7 . The method of claim 1 , wherein the object detection machine learning model includes a feature encoder, a region proposal network (RPN), and a region of interest (ROI) pooling layer.
8 . A system comprising one or more processors and memory storing instructions that, in response to execution by the one or more processors, cause the one or more processors to:
perform weak augmentation on images from a target domain (DT) to generate a first set of weakly augmented target domain images;
perform strong augmentation on images from a source domain (DS) and the images from the DT to generate a second set of strongly augmented images;
process the second set of strongly augmented images to generate a third set of inter- domain mixes of the images from the DS and the DT; and
jointly train an object detection machine learning model, as a student machine learning model, with a teacher machine learning model using the first and third sets.
9 . The system of claim 8 , further including instructions to process images from the second set to generate a fourth set of intra-domain mixes of the images from the DS.
10 . The system of claim 9 , wherein the object detection machine learning model is further trained using the fourth set.
11 . The system of claim 8 , wherein the instructions to jointly train include instructions to update weights of the teacher machine learning model based on an exponential moving average of weights of the object detection machine learning model.
12 . The system of claim 8 , wherein output generated using the teacher machine learning model includes pseudo labels that are used to train the object detection machine learning model.
13 . The system of claim 8 , further including instructions to perform adversarial training of the object detection machine learning model and a discriminator machine learning model, wherein to perform the adversarial training includes instructions to apply the discriminator machine learning model to a feature embedding generated by the object detection machine learning model to generate a predicted classification of the feature embedding as being in the DS or the DT.
14 . The system of claim 8 , wherein the object detection machine learning model includes a feature encoder, a region proposal network (RPN), and a region of interest (ROI) pooling layer.
15 . A method for performing cross-domain training of an object detection machine learning model, the method implemented using one or more processors, comprising:
performing strong augmentation on images from a source domain (DS) and images from a target domain (DT) to generate a first set of strongly augmented source domain images and a second set of strongly augmented target domain images;
mixing images from the first set of strongly augmented source domain images with images from the second set of strongly augmented target domain images to generate a third set of inter-domain mixed images of the images from the DS and the DT;
processing images from the first or second sets to generate a fourth set of intra-domain mixed images from the DS or the DT; and
jointly training the object detection machine learning model, as a student machine learning model, with a teacher machine learning model using the third and fourth sets.
16 . The method of claim 15 , further including performing weak augmentation on images from the DT to generate a fifth set of weakly augmented target domain images.
17 . The method of claim 16 , wherein the object detection machine learning model is further trained using the fifth set.
18 . The method of claim 15 , wherein the jointly training includes updating weights of the teacher machine learning model based on an exponential moving average of weights of the object detection machine learning model.
19 . The method of claim 15 , wherein output generated using the teacher machine learning model includes pseudo labels that are used to train the object detection machine learning model.
20 . The method of claim 15 , further including performing adversarial training of the object detection machine learning model and a discriminator machine learning model, wherein performing the adversarial training includes applying the discriminator machine learning model to a feature embedding generated by the object detection machine learning model to generate a predicted classification of the feature embedding as being in the source DS or the DT.