Dense and discriminative neural network architectures for improved object detection and instance segmentation
This disclosure relates to improved techniques for performing computer vision functions, including common object detection and instance segmentation. The techniques described herein utilize neural network architectures to perform these functions in various types of images, such as natural images, UAV images, satellite images, and other images. The neural network architecture can include a dense location regression network that performs object localization and segmentation functions, at least in part, by generating offset information for multiple sub-regions of candidate object proposals, and utilizing this dense offset information to derive final predictions for locations of target objects. The neural network architecture also can include a discriminative region-of-interest (RoI) pooling network that performs classification of the localized objects, at least in part, by sampling various sub-regions of candidate proposals and performing adaptive weighting to obtain discriminative features.
1. A computer vision system comprising:
one or more computing devices comprising one or more processors and one or more non-transitory storage devices for storing instructions, wherein execution of the instructions by the one or more processors causes the one or more computing devices to:
receive an image comprising an object at a neural network architecture;
identify, using a region proposal network of the neural network architecture, at least one proposal corresponding to the object in the image;
segment the at least one proposal into a plurality of sub-regions, a first group of sub-regions are predicted to belong to the object in the image and a second group of sub-regions are predicted to belong to a background portion of the image;
generate, using a dense local regression network of the neural network architecture, separate offset information based on each of the plurality of sub-regions, wherein the offset information for each sub-region comprises a separate prediction for localizing the object in the image, the dense local regression network is configured to generate binary overlap predictions that predict whether the plurality of sub-regions belong to the object in the image or the background portion of the image; and
generate, using the dense local regression network of the neural network architecture, a final localization prediction for the object in the image based, at least in part, on the offset information generated for the plurality of sub-regions, the offset information utilized to generate the final localization prediction only includes the offset information associated with the first group of sub-regions and does not include the offset information associated with the second group of sub-regions.
2. The computer vision system of claim 1 , wherein generating the final localization prediction for the object in the image includes averaging the offset information from the first group of sub-regions.
3. The computer vision system of claim 1 , wherein:
the neural network architecture includes a discriminative region-of-interest (RoI) pooling network; and
the discriminative RoI pooling network is configured to classify the at least one proposal in the image.
4. The computer vision system of claim 3 , wherein classifying the at least one proposal in the image includes:
segmenting the at least one proposal into a second group of sub-regions; executing an offset prediction function to predict offsets corresponding to the second group of sub-regions;
analyzing features included the second group of sub-regions; and
executing an adaptive weight pooling function to assign weights to the features.
5. The computer vision system of claim 4 , wherein classifying the at least one proposal in the image further includes:
generating a classification score for the at least one proposal indicating a probability that the at least one proposal includes an object associated with a particular class or label.
6. The computer vision system of claim 1 , wherein:
the offset information generated by the dense local regression network includes separate predictions for a location of a bounding box to localize the object in the image; and
the final localization prediction includes a final prediction location of the bounding box.
7. The computer vision system of claim 1 , wherein:
the offset information generated by the dense local regression network includes separate predictions for a location of an instance mask to localize the object in the image; and
the final localization prediction includes a final prediction location of the instance mask utilized to identify the object.
8. The computer vision system of claim 1 , wherein the computer vision system is integrated into, or communicates with, one or more of: a surveillance system, a satellite imaging system, an unmanned aerial vehicle system, an autonomous vehicle system, a traffic control system, or an image editing application.
9. A method implemented via execution of computing instructions configured to run at one or more processors and configured to be stored at non-transitory computer-readable media, the method comprising:
receiving an image at a neural network architecture;
identifying, using a region proposal network of the neural network architecture, at least one proposal corresponding to an object in the image;
segmenting the at least one proposal into a plurality of sub-regions, a first group of sub-regions are predicted to belong to the object in the image and a second group of sub-regions are predicted to belong to a background portion of the image;
generating, using a dense local regression network of the neural network architecture, separate offset information based on each of the plurality of sub-regions, wherein the offset information for each sub-region comprises a separate prediction for localizing the object in the image, the dense local regression network is configured to generate binary overlap predictions that predict whether the plurality of sub-regions belong to the object in the image or the background portion of the image; and
generating, using the dense local regression network of the neural network architecture, a final localization prediction for the object in the image based, at least in part, on the offset information generated for the plurality of sub-regions, the offset information utilized to generate the final localization prediction only includes the offset information associated with the first group of sub-regions and does not include offset information associated with the second group of sub-regions.
10. The method of claim 9 , wherein generating the final localization prediction for the object in the image includes averaging the offset information from the first group of sub-regions.
11. The method of claim 9 , wherein:
the neural network architecture includes a discriminative region-of-interest (RoI) pooling network; and
the discriminative RoI pooling network is configured to classify the at least one proposal in the image.
12. The method of claim 11 , wherein classifying the at least one proposal in the image includes:
segmenting the at least one proposal into a second group of sub-regions;
executing an offset prediction function to predict offsets corresponding to the second group of sub-regions;
analyzing features included the second group of sub-regions; and
executing an adaptive weight pooling function to assign weights to the features.
13. The method of claim 12 , wherein classifying the at least one proposal in the image further includes:
generating a classification score for the at least one proposal indicating a probability that the at least one proposal includes an object associated with a particular class or label.
14. The method of claim 9 , wherein:
the offset information generated by the dense local regression network includes separate predictions for a location of a bounding box to localize the object in the image; and
the final localization prediction includes a final prediction location of the bounding box.
15. The method of claim 9 , wherein:
the offset information generated by the dense local regression network includes separate predictions for a location of an instance mask to localize the object in the image; and
the final localization prediction includes a final prediction location of the instance mask utilized to identify the object.
16. A computer program product comprising a non-transitory computer-readable medium including instructions for causing a computer to:
receive an image comprising an object at a neural network architecture;
identify, using a region proposal network of the neural network architecture, at least one proposal corresponding to the object in the image;
segment the at least one proposal into a plurality of sub-regions, a first group of sub-regions are predicted to belong to the object in the image and a second group of sub-regions are predicted to belong to a background portion of the image;
generate, using a dense local regression network of the neural network architecture, separate offset information based on each of the plurality of sub-regions, wherein the offset information for each sub-region comprises a separate prediction for localizing the object in the image, the dense local regression network is configured to generate binary overlap predictions that predict whether the plurality of sub-regions belong to the object in the image or the background portion of the image; and
generate, using the dense local regression network of the neural network architecture, a final localization prediction for the object in the image based, at least in part, on the offset information generated for the plurality of sub-regions, the offset information utilized to generate the final localization prediction only includes the offset information associated with the first group of sub-regions and does not include offset information associated with the second group of sub-regions.
17. A computer vision system comprising:
one or more computing devices comprising one or more processors and one or more non-transitory storage devices for storing instructions, wherein execution of the instructions by the one or more processors causes the one or more computing devices to:
receive an image comprising an object at a neural network architecture;
identify, using a region proposal network of the neural network architecture, at least one proposal corresponding to the object in the image, the neural network architecture includes a discriminative region-of-interest (RoI) pooling network, the discriminative RoI pooling network is configured to classify the at least one proposal in the image by:
segmenting the at least one proposal into a second group of sub-regions;
executing an offset prediction function to predict offsets corresponding to the second group of sub-regions;
analyzing features included in the second group of sub-regions; and
executing an adaptive weight pooling function to assign weights to the features;
segment the at least one proposal into a plurality of sub-regions;
generate, using a dense local regression network of the neural network architecture, separate offset information based on each of the plurality of sub-regions, wherein the offset information for each sub-region comprises a separate prediction for localizing the object in the image; and
generate, using the dense local regression network of the neural network architecture, a final localization prediction for the object in the image based, at least in part, on the offset information generated for the plurality of sub-regions.
18. The computer vision system of claim 17 , wherein classifying the at least one proposal in the image further includes:
generating a classification score for the at least one proposal indicating a probability that the at least one proposal includes an object associated with a particular class or label.
19. A method implemented via execution of computing instructions configured to run at one or more processors and configured to be stored at non-transitory computer-readable media, the method comprising:
receiving an image at a neural network architecture;
identifying, using a region proposal network of the neural network architecture, at least one proposal corresponding to an object in the image, the neural network architecture includes a discriminative region-of-interest (RoI) pooling network, the discriminative RoI pooling network is configured to classify the at least one proposal in the image by:
segmenting the at least one proposal into a second group of sub- regions;
executing an offset prediction function to predict offsets corresponding to the second group of sub-regions;
analyzing features included in the second group of sub-regions; and
executing an adaptive weight pooling function to assign weights to the features;
segmenting the at least one proposal into a plurality of sub-regions;
generating, using a dense local regression network of the neural network architecture, separate offset information based on each of the plurality of sub-regions, wherein the offset information for each sub-region comprises a separate prediction for localizing the object in the image; and
generating, using the dense local regression network of the neural network architecture, a final localization prediction for the object in the image based, at least in part, on the offset information generated for the plurality of sub-regions.
20. The method of claim 19 , wherein classifying the at least one proposal in the image further includes:
generating a classification score for the at least one proposal indicating a probability that the at least one proposal includes an object associated with a particular class or label.