Image grounding with modularized graph attentive networks
A system may include a memory and a processor in communication with the memory. The processor may be configured to perform operations. The operations may include receiving an input, extracting features from the input, and mining object relations using the features. The operations may include determining feature vectors using the object relations and generating, using the feature vectors, an output indicating a target region, wherein the target region corresponds to the input.
1. A system, said system comprising:
a memory; and
a processor in communication with said memory, said processor being configured to perform operations, said operations comprising:
receiving an input;
extracting features from said input, wherein extracting features from said input comprises:
using an attention network to extract textual features for a plurality of specific modules, the plurality of specific modules including at least a subject module, a location module, and a relation module, wherein the attention network parses different components of the input for each specific module, including parsing a subject component for the subject module, a location component for the location module, and a relation component for the relation module;
mining object relations using said features;
determining feature vectors using said object relations; and
generating, using said feature vectors, an output indicating a target region, wherein said target region corresponds to said input.
2. The system of claim 1 , wherein:
said input includes a textual statement with a plurality of textual modules.
3. The system of claim 1 , said operations further comprising:
obtaining visual features for a plurality of regions in said input by performing message passing.
4. The system of claim 1 , said operations further comprising:
applying a text-guided graph attentive neural network to a module of said input.
5. The system of claim 1 , said operations further comprising:
determining visual regional node aggregate context information based on neighborhood nodes.
6. The system of claim 1 , said operations further comprising:
matching, bilinearly, modularized similarity between modularized textual features of said input and modularized visual features of said input.
7. A computer-implemented method, said method comprising:
receiving an input;
extracting features from said input, wherein extracting features from said input comprises:
using an attention network to extract textual features for a plurality of specific modules, the plurality of specific modules including at least a subject module, a location module, and a relation module, wherein the attention network parses different components of the input for each specific module, including parsing a subject component for the subject module, a location component for the location module, and a relation component for the relation module;
mining object relations using said features;
determining feature vectors using said object relations; and
generating, using said feature vectors, an output indicating a target region, wherein said target region corresponds to said input.
8. The computer-implemented method of claim 7 , wherein:
said input includes a textual statement with a plurality of textual modules.
9. The computer-implemented method of claim 8 , further comprising:
obtaining visual features for a plurality of regions in said input by performing message passing.
10. The computer-implemented method of claim 7 , further comprising:
applying a text-guided graph attentive neural network to a module of said input.
11. The computer-implemented method of claim 7 , further comprising:
determining visual regional node aggregate context information based on neighborhood nodes.
12. The computer-implemented method of claim 7 , further comprising:
matching, bilinearly, modularized similarity between modularized textual features of said input and modularized visual features of said input.
13. The computer-implemented method of claim 7 , further comprising:
aggregating, linearly, similarities over a plurality of modules to obtain a final similarity, wherein said final similarity is used in generating said output.
14. A computer program product, said computer program product comprising a computer readable storage medium having program instructions embodied therewith, said program instructions executable by a processor to cause said processor to perform a function, said function comprising:
receiving an input;
extracting features from said input, wherein extracting features from said input comprises:
using an attention network to extract textual features for a plurality of specific modules, the plurality of specific modules including at least a subject module, a location module, and a relation module, wherein the attention network parses different components of the input for each specific module, including parsing a subject component for the subject module, a location component for the location module, and a relation component for the relation module;
mining object relations using said features;
determining feature vectors using said object relations; and
generating, using said feature vectors, an output indicating a target region, wherein said target region corresponds to said input.
15. The computer program product of claim 14 , said function further comprising:
applying a text-guided graph attentive neural network to a module of said input.
16. The computer program product of claim 14 , said function further comprising:
determining visual regional node aggregate context information based on neighborhood nodes.
17. The computer program product of claim 14 , said function further comprising:
matching, bilinearly, modularized similarity between modularized textual features of said input and modularized visual features of said input.