Domain adaptation for instance detection and segmentation

1. A method for domain adaptation, comprising:

aligning image level features between a source domain and a target domain based on an adversarial learning process while training a domain discriminator;

selecting, using the domain discriminator, unlabeled samples from the target domain that are furthest away from existing annotated samples from the target domain;

selecting, by a processor device, based on a prediction score of each of the unlabeled samples, samples with lower prediction scores; and

annotating the samples with the lower prediction scores.

2. The method as recited in claim 1 , further comprising:

iteratively retraining a model that annotates the unlabeled samples based on the annotated samples with the lower prediction scores, wherein the model implements at least one predetermined task.

3. The method as recited in claim 2 , wherein the at least one predetermined task includes at least one of instance object detection and segmentation.

4. The method as recited in claim 2 , wherein retraining the model further comprises:

inputting an updated label set including the annotated samples with the lower prediction scores into an image-level convolutional neural network (CNN) to generate at least one feature;

based on the at least one feature, propagating the updated label set to a region of interest level (ROI-level) CNN; and

generating output bounding boxes as at least one object detection.

5. The method as recited in claim 4 , further comprising:

predicting an instance segmentation map within each bounding box.

6. The method as recited in claim 1 , wherein aligning the image level features between the source domain and the target domain based on the adversarial learning process further comprises:

applying an adversarial loss function to encourage a distribution of labeled samples and the unlabeled samples from a label set;

selecting, by the processor device, at least one higher diversity score unlabeled sample from the unlabeled samples; and

selecting at least one lower prediction score higher diversity score unlabeled sample from the at least one higher diversity score unlabeled sample.

7. The method as recited in claim 6 , further comprising:

annotating the at least one lower prediction score higher diversity score unlabeled sample; and

updating the label set with at least one annotated lower prediction score higher diversity score unlabeled sample to form an updated labeled set.

8. The method as recited in claim 6 , wherein selecting the at least one lower prediction score higher diversity score unlabeled sample from the unlabeled samples further comprises:

using prediction scores of the unlabeled samples as confidence scores.

9. The method as recited in claim 1 , wherein the source domain and the target domain are selected from at least one of different geographical areas, different weather conditions and different lighting conditions.

10. The method as recited in claim 1 , wherein selecting the at least one higher diversity score unlabeled sample from the unlabeled samples further comprises:

selecting unlabeled images that are furthest away from existing annotated images in the label set.

11. The method as recited in claim 1 , further comprising:

using a supervised loss function and ground truth labels from the source domain and the target domain to train at least one image-level convolutional neural network (CNN).

12. A computer system for domain adaptation, comprising:

a processor device operatively coupled to a memory device, the processor device being configured to:

align image level features between a source domain and a target domain based on an adversarial learning process while training a domain discriminator;

select, using the domain discriminator, unlabeled samples from the target domain that are far away from existing annotated samples from the target domain;

select based on a prediction score of each of the unlabeled samples, samples with lower prediction scores; and

annotate the samples with the lower prediction scores.

13. The system as recited in claim 12 , wherein the processor device is further configured to:

iteratively retrain a model that annotates the unlabeled samples based on the annotated samples with the lower prediction scores, wherein the model implements at least one predetermined task.

14. The system as recited in claim 13 , wherein the at least one predetermined task includes at least one of instance object detection and segmentation.

15. The system as recited in claim 13 , wherein, when retraining the model, the processor device is further configured to:

input an updated label set including the annotated samples with the lower prediction scores into an image-level convolutional neural network (CNN) to generate at least one feature;

based on the at least one feature, propagate the updated label set to a region of interest level (ROI-level) CNN; and

generate output bounding boxes as at least one object detection.

16. The system as recited in claim 15 , wherein the processor device is further configured to:

predict an instance segmentation map within each bounding box.

17. The system as recited in claim 13 , wherein, when aligning the image level features between the source domain and the target domain based on the adversarial learning process, the processor device is further configured to:

apply an adversarial loss function to encourage a distribution of labeled samples and the unlabeled samples from a label set;

select at least one higher diversity score unlabeled sample from the unlabeled samples; and

selecting at least one lower prediction score higher diversity score unlabeled sample from the at least one higher diversity score unlabeled sample.

18. The system as recited in claim 12 , wherein the source domain and the target domain are selected from at least one of different geographical areas, different weather conditions and different lighting conditions.

19. The system as recited in claim 12 , wherein the processor device is further configured to:

use a supervised loss function and ground truth labels from the source domain and the target domain to train at least one image-level convolutional neural network (CNN).

20. A computer program product for domain adaptation, the computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computing device to cause the computing device to perform the method comprising:

aligning image level features between a source domain and a target domain based on an adversarial learning process while training a domain discriminator;

selecting, using the domain discriminator, unlabeled samples from the target domain that are far away from existing annotated samples from the target domain;

selecting, by a processor device, based on a prediction score of each of the unlabeled samples, samples with lower prediction scores; and

annotating the samples with the lower prediction scores.