Human parsing techniques utilizing neural network architectures
This disclosure relates to improved techniques for performing human parsing functions using neural network architectures. The neural network architecture can model human objects in images using a hierarchal graph of interconnected nodes that correspond to anatomical features at various levels. Multi-level inference information can be generated for each of the nodes using separate inference processes. The multi-level inference information for each node can be combined or fused to generate final predictions for each of the nodes. Parsing results may be generated based on the final predictions.
1. A 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, at a neural network architecture comprising a human parsing network, an image comprising at least one human object;
utilize a hierarchal graph comprising a plurality of nodes to model the at least one human object, wherein the nodes correspond to anatomical features associated with a human body;
generate inference information for the nodes in the hierarchal graph, wherein generating inference information includes:
deriving, with the neural network architecture, direct inference information for each of the nodes included in the hierarchal graph;
deriving, with the neural network architecture, top-down inference information for at least a portion of the nodes included in the hierarchal graph; and
deriving, with the neural network architecture, bottom-up inference information for at least a portion of the nodes included in the hierarchal graph; and
generate, with the neural network architecture, parsing results based, at least in part, on the inference information associated with the nodes.
2. The system of claim 1 , wherein:
the neural network architecture comprises a fusion network; and
for each node included in the hierarchal graph, the fusion network combines or fuses the inference information derived for the node to obtain a final prediction for the node.
3. The system of claim 2 , wherein:
the fusion network includes a gating function that is configured to compute confidence scores associated with the direct inference information, the top-down inference information, and the bottom-up inference information; and
the confidence scores are utilized to determine an extent to which the direct inference information, the top-down inference information, and the bottom-up inference information are utilized to generate the final prediction for the node.
4. The system of claim 2 , wherein final predictions for the nodes are combined to generate the parsing results for the at least one human object.
5. The system of claim 1 , wherein the hierarchal graph utilized to model the at least one human object comprises:
a top tier node that corresponds to an entirety of the at least one human object;
second tier nodes that correspond to an upper body portion and lower body portion of the at least one human object; and
third tier nodes that correspond to anatomical subcomponents of the upper body portion and the lower body portion of the at least one human object.
6. The system of claim 1 , wherein:
the neural network architecture comprises a direct inference network; and
the direct inference network is configured to generate the direct inference information directly from raw data associated with the image.
7. The system of claim 1 , wherein:
the neural network architecture comprises a top-down inference network; and
the top-down inference network is configured to generate the top-down inference information for a node based, at least in part, on the direct inference information derived for the node and contextual information from one or more parent nodes included in the hierarchical graph.
8. The system of claim 1 , wherein:
the neural network architecture comprises a bottom-up inference network; and
the bottom-up inference network is configured to generate the bottom-up inference information for a node based, at least in part, on the direct inference information derived for the node and contextual information from one or more child nodes included in the hierarchical graph.
9. The system of claim 1 , wherein:
the direct inference information, the top-down inference information, and the bottom-up inference information each comprise one or more segmentation maps, and the one or more segmentation maps are combined or fused to generate final predictions for each of the nodes.
10. The system of claim 1 , wherein the human parsing network is integrated with, or communicates with, one or more of: a surveillance application, a facial recognition application, a fashion-related application, or a social media application.
11. A method comprising:
receiving, at a neural network architecture comprising a human parsing network, an image comprising at least one human object;
utilizing a hierarchal graph comprising a plurality of nodes to model the at least one human object, wherein the nodes correspond to anatomical features associated with a human body;
generating inference information for the nodes in the hierarchal graph, wherein generating inference information includes:
deriving, with the neural network architecture, direct inference information for at least a portion of the nodes included in the hierarchal graph;
deriving, with the neural network architecture, top-down inference information for at least a portion of the nodes included in the hierarchal graph; and
deriving, with the neural network architecture, bottom-up inference information for at least a portion of the nodes included in the hierarchal graph; and
generating, with the neural network architecture, parsing results based, at least in part, on the inference information associated with the nodes.
12. The method of claim 11 , wherein:
the neural network architecture comprises a fusion network; and
for each node included in the hierarchal graph, the fusion network combines or fuses the inference information derived for the node to obtain a final prediction for the node.
13. The method of claim 12 , wherein:
the fusion network includes a gating function that is configured to compute confidence scores associated with the direct inference information, the top-down inference information, and the bottom-up inference information; and
the confidence scores are utilized to determine an extent to which the direct inference information, the top-down inference information, and the bottom-up inference information are utilized to generate the final prediction for the node.
14. The method of claim 12 , wherein final predictions for the nodes are combined to generate the parsing results for the at least one human object.
15. The method of claim 11 , wherein the hierarchal graph utilized to model the at least one human object comprises:
a top tier node that corresponds to an entirety of the at least one human object;
second tier nodes that correspond to an upper body portion and lower body portion of the at least one human object; and
third tier nodes that correspond to sub-features of the upper body portion and the lower body portion of the at least one human object.
16. The method of claim 11 , wherein:
the neural network architecture comprises a direct inference network; and
the direct inference network is configured to generate the direct inference information directly from raw data associated with the image.
17. The method of claim 11 , wherein:
the neural network architecture comprises a top-down inference network; and
the top-down inference network is configured to generate the top-down inference information for a node based, at least in part, on the direct inference information derived for the node and contextual information from one or more parent nodes included in the hierarchical graph.
18. The method of claim 11 , wherein:
the neural network architecture comprises a bottom-up inference network; and
the bottom-up inference network is configured to generate the bottom-up inference information for a node based, at least in part, on the direct inference information derived for the node and contextual information from one or more child nodes included in the hierarchical graph.
19. The method of claim 11 , wherein:
the direct inference information, the top-down inference information, and the bottom-up inference information each comprise one or more segmentation maps, and the one or more segmentation maps are combined or fused to generate final predictions for each of the nodes.
20. A computer program product comprising a non-transitory computer-readable medium, including instructions for causing a computer to:
receive, at a neural network architecture comprising a human parsing network, an image comprising at least one human object;
utilize a hierarchal graph comprising a plurality of nodes to model the at least one human object, wherein the nodes correspond to anatomical features associated with a human body;
generate inference information for the nodes in the hierarchal graph, wherein generating inference information includes:
deriving, with the neural network architecture, direct inference information for at least a portion of the nodes included in the hierarchal graph;
deriving, with the neural network architecture, top-down inference information for at least a portion of the nodes included in the hierarchal graph; and
deriving, with the neural network architecture, bottom-up inference information for at least a portion of the nodes included in the hierarchal graph; and
generate, with the neural network architecture, parsing results based, at least in part, on the inference information associated with the nodes.