Generating improved panoptic segmented digital images based on panoptic segmentation neural networks that utilize exemplar unknown object classes
This disclosure describes one or more implementations of a panoptic segmentation system that generates panoptic segmented digital images that classify both known and unknown instances of digital images. For example, the panoptic segmentation system builds and utilizes a panoptic segmentation neural network to discover, cluster, and segment new unknown object subclasses for previously unknown object instances. In addition, the panoptic segmentation system can determine additional unknown object instances from additional digital images. Moreover, in some implementations, the panoptic segmentation system utilizes the newly generated unknown object subclasses to refine and tune the panoptic segmentation neural network to improve the detection of unknown object instances in input digital images.
1. A non-transitory computer-readable medium storing instructions that, when executed by at least one processor, cause a computing device to:
determine a subset of unknown object instances portrayed within a set of digital images utilizing a panoptic segmentation neural network that generates a plurality of object feature vectors for the subset of unknown object instances;
determine a first unknown object subclass and a second unknown object subclass by grouping the subset of unknown object instances utilizing an object feature clustering algorithm and the plurality of object feature vectors for the subset of unknown object instances;
generate an additional object feature vector corresponding to an additional unknown object instance identified from an additional digital image utilizing the panoptic segmentation neural network; and
add the additional unknown object instance to the first unknown object subclass based on comparing the additional object feature vector to an exemplar object feature vector corresponding to an exemplar object instance from the first unknown object subclass.
2. The non-transitory computer-readable medium of claim 1 , further comprising instructions that, when executed by the at least one processor, cause the computing device to determine the subset of unknown object instances by:
determining regional proposals for a plurality of objects in the set of digital images utilizing the panoptic segmentation neural network; and
classifying the regional proposals utilizing the panoptic segmentation neural network to an unknown object class to determine the subset of unknown object instances.
3. The non-transitory computer-readable medium of claim 1 , further comprising instructions that, when executed by the at least one processor, cause the computing device to determine the first unknown object subclass by:
generating a plurality of object feature clusters comprising a first object feature cluster associated with the first unknown object subclass; and
determining a combined cosine distance between a centroid of the first object feature cluster and object feature vectors clustered in the first object feature cluster.
4. The non-transitory computer-readable medium of claim 3 , further comprising instructions that, when executed by the at least one processor, cause the computing device to filter out object feature vectors from the first object feature cluster that are associated with unknown object instances originating from a same digital image within the set of digital images.
5. The non-transitory computer-readable medium of claim 1 , further comprising instructions that, when executed by the at least one processor, cause the computing device to refine parameters of the panoptic segmentation neural network utilizing the first unknown object subclass and the second unknown object subclass.
6. The non-transitory computer-readable medium of claim 1 , further comprising instructions that, when executed by the at least one processor, cause the computing device to:
receive an input digital image comprising a plurality of object instances;
segment the input digital image utilizing the panoptic segmentation neural network to classify a first object instance of the plurality of object instances to a known class; and
segment the input digital image utilizing the panoptic segmentation neural network to segment the input digital image utilizing the panoptic segmentation neural network to classify a second object instance of the plurality of object instances to the first unknown object subclass.
7. The non-transitory computer-readable medium of claim 1 , further comprising instructions that, when executed by the at least one processor, cause the computing device to generate the additional object feature vector by:
generating, utilizing the panoptic segmentation neural network, a plurality of object feature vectors for a subsequent batch of digital images comprising the additional digital image;
determining that additional unknown object instances, which correspond to a subset of the plurality of object feature vectors for the subsequent batch of digital images, correspond to exemplar object instances from the first unknown object subclass; and
adding the additional unknown object instances to the first unknown object subclass based on the correspondence.
8. The non-transitory computer-readable medium of claim 1 , further comprising instructions that, when executed by the at least one processor, cause the computing device to add the additional unknown object instance to the first unknown object subclass by:
determining a feature similarity score between the additional object feature vector and object feature vectors from exemplar object instances in the first unknown object subclass; and
based on the feature similarity score being greater than an unknown object subclass similarity threshold, adding the additional unknown object instance to the first unknown object subclass as a new exemplar object instance.
9. A system comprising:
one or more memory devices; and
one or more server devices configured to cause the system to:
determine a subset of unknown object instances portrayed within a set of digital images utilizing a panoptic segmentation neural network that generates a plurality of object feature vectors for the subset of unknown object instances;
determine a first unknown object subclass and a second unknown object subclass by grouping the subset of unknown object instances utilizing an object feature clustering algorithm and the plurality of object feature vectors for the subset of unknown object instances;
generate an additional object feature vector corresponding to an additional unknown object instance identified from an additional digital image utilizing the panoptic segmentation neural network; and
add the additional unknown object instance to the first unknown object subclass based on comparing the additional object feature vector to an exemplar object feature vector corresponding to an exemplar object instance from the first unknown object subclass.
10. The system of claim 9 , wherein the one or more server devices are further configured to cause the system to determine the subset of unknown object instances by:
determining regional proposals for a plurality of objects in the set of digital images; and
classifying the regional proposals utilizing the panoptic segmentation neural network to an unknown object class.
11. The system of claim 10 , wherein the one or more server devices are further configured to cause the system to determine the first unknown object subclass by:
generating a plurality of object feature clusters comprising a first object feature cluster associated with the first unknown object subclass; and
determining a distance between a centroid of the first object feature cluster and object feature vectors clustered in the first object feature cluster.
12. The system of claim 11 , wherein the one or more server devices are further configured to cause the system to filter out object feature vectors from the first object feature cluster that are associated with unknown object instances originating from a same digital image within the set of digital images.
13. The system of claim 10 , wherein the one or more server devices are further configured to cause the system to refine parameters of the panoptic segmentation neural network utilizing the first unknown object subclass and the second unknown object subclass.
14. The system of claim 10 , wherein the one or more server devices are further configured to build an object segmentation neural network utilizing the first unknown object subclass as a ground truth to classify object instances within digital images.
15. The system of claim 14 , wherein the one or more server devices are further configured to cause the system to:
convert the first unknown object subclass to a new known object class associated with a label based on determining the label for the first unknown object subclass; and
build the object segmentation neural network to segment object instances utilizing the new known object class with the label.
16. A computer-implemented method comprising:
determining a subset of unknown object instances portrayed within a set of digital images utilizing a panoptic segmentation neural network that generates a plurality of object feature vectors for the subset of unknown object instances;
determining a first unknown object subclass and a second unknown object subclass by grouping the subset of unknown object instances utilizing an object feature clustering algorithm and the plurality of object feature vectors for the subset of unknown object instances;
generating an additional object feature vector corresponding to an additional unknown object instance identified from an additional digital image utilizing the panoptic segmentation neural network; and
adding the additional unknown object instance to the first unknown object subclass based on comparing the additional object feature vector to an exemplar object feature vector corresponding to an exemplar object instance from the first unknown object subclass.
17. The computer-implemented method of claim 16 , further comprising determining the subset of unknown object instances by:
determining regional proposals for a plurality of objects in the set of digital images utilizing the panoptic segmentation neural network; and
classifying the regional proposals utilizing the panoptic segmentation neural network to an unknown object class.
18. The computer-implemented method of claim 16 , further comprising determining the first unknown object subclass by:
generating a plurality of object feature clusters comprising a first object feature cluster associated with the first unknown object subclass; and
determining a distance between a centroid of the first object feature cluster and object feature vectors clustered in the first object feature cluster.
19. The computer-implemented method of claim 16 , further comprising refining parameters of the panoptic segmentation neural network utilizing the first unknown object subclass and the second unknown object subclass.
20. The computer-implemented method of claim 16 , further comprising building an object segmentation neural network utilizing the first unknown object subclass as a ground truth to classify object instances within digital images.