Object detection based on object relation
Various embodiments described herein provide for detection of a particular object within a scene depicted by image data by using a coarse-to-fine approach/strategy based on one or more relationships of objects depicted within the scene.
1. A method comprising:
accessing, by one or more computers, image data that comprises an initial image, the initial image depicting a scene to be searched for a target object; and
processing, by the one or more computers, the scene to search for the target object by at least:
detecting, by processing the scene depicted in the initial image, one or more first anchor objects depicted in the scene that are each related to the target object; and
determining a set of regions of the scene by identifying, for each particular anchor object in the one or more first anchor objects, a region of the scene that is relative to the particular anchor object and has a first resolution; and
for a particular region in the set of regions, searching for the target object in the particular region at a second resolution, the second resolution being one of higher than or equal to the first resolution of the particular region at which the particular region was identified.
2. The method of claim 1 , wherein searching for the target object in the particular region comprises:
searching a first region from the set of regions for the target object;
in response to determining that the target object is not depicted in the first region using a result of the search, searching a second different region from the set of regions for the target object.
3. The method of claim 1 , comprising:
generating the initial image by down sampling an original image to the first resolution, the original image depicting the scene.
4. The method of claim 3 , wherein searching for the target object in the particular region at the second resolution comprises:
extracting the particular region from the original image;
generating a particular region image by down sampling the extracted particular region to the second resolution; and
searching for the target object in the particular region image.
5. The method of claim 1 , wherein detecting the one or more first anchor objects depicted in the scene using a process with a first detection rate of detecting the one or more first anchor objects in the initial image that is higher than a second detection rate of detecting the target object in the initial image.
6. The method of claim 1 , wherein for the particular anchor object, the identifying the region of the scene relative to the particular anchor object comprises:
determining a bounding box relative to the particular anchor object, the bounding box defining the region.
7. The method of claim 1 , wherein for the particular anchor object, the identifying the region of the scene relative to the particular anchor object comprises:
processing, by a machine learning model for region predication, the scene to identify the region.
8. The method of claim 1 , wherein the one or more first anchor objects comprise a human individual.
9. The method of claim 8 , wherein searching for the target object in the particular region at the second resolution comprises:
determining whether to search for a second anchor object in the particular region, the second anchor object relating to the target object; and
in response to determining to search for the second anchor object in the particular region:
searching, by processing the particular region at the second resolution, for a depiction of the second anchor object in the particular region; and
searching for the target object in the particular region using a result of whether searching the particular region at the second resolution for the depiction of the second anchor object caused detection of a second set of anchor objects in the particular region, the second anchor object comprising at least one of a human hand or a human head.
10. The method of claim 1 , wherein the one or more first anchor objects comprise a vehicle.
11. The method of claim 1 , wherein searching for the target object in the particular region at the second resolution comprises:
determining whether to search for a second anchor object in the particular region, the second anchor object relating to the target object; and
in response to determining to search for the second anchor object in the particular region:
searching, by processing the particular region at the second resolution, for a depiction of the second anchor object in the particular region; and
searching for the target object in the particular region using a result of whether searching the particular region at the second resolution for the depiction of the second anchor object caused detection of a second set of anchor objects in the particular region.
12. One or more non-transitory computer storage media encoded with instructions that, when executed by one or more computers, cause the one or more computers to perform operations comprising:
accessing image data that comprises an initial image, the initial image depicting a scene to be searched for a target object; and
processing the scene to search for the target object by at least:
detecting, by processing the scene depicted in the initial image, one or more first anchor objects depicted in the scene that are each related to the target object; and
determining a set of regions of the scene by identifying, for each particular anchor object in the one or more first anchor objects, a region of the scene that is relative to the particular anchor object and has a first resolution; and
for a particular region in the set of regions, searching for the target object in the particular region at a second resolution, the second resolution being one of higher than or equal to the first resolution of the particular region at which the particular region was identified.
13. The one or more non-transitory computer storage media of claim 12 , wherein searching for the target object in the particular region comprises:
searching a first region from the set of regions for the target object;
in response to determining that the target object is not depicted in the first region using a result of the search, searching a second different region from the set of regions for the target object.
14. The one or more non-transitory computer storage media of claim 12 , wherein the operations comprise:
generating the initial image by down sampling an original image to the first resolution, the original image depicting the scene.
15. The one or more non-transitory computer storage media of claim 14 , wherein searching for the target object in the particular region at the second resolution comprises:
extracting the particular region from the original image;
generating a particular region image by down sampling the extracted particular region to the second resolution; and
searching for the target object in the particular region image.
16. The one or more non-transitory computer storage media of claim 12 , wherein the one or more first anchor objects comprise a human individual.
17. The one or more non-transitory computer storage media of claim 16 , wherein searching for the target object in the particular region at the second resolution comprises:
determining whether to search for a second anchor object in the particular region, the second anchor object relating to the target object; and
in response to determining to search for the second anchor object is to be detected for in the particular region:
searching, by processing the particular region at the second resolution, for a depiction of the second anchor object in the particular region; and
searching for the target object in the particular region using a result of whether searching the particular region at the second resolution for the depiction of the second anchor object caused detection of a second set of anchor objects in the particular region, the second anchor object comprising at least one of a human hand or a human head.
18. The one or more non-transitory computer storage media of claim 12 , wherein the one or more first anchor objects comprise a vehicle.
19. The one or more non-transitory computer storage media of claim 12 , wherein searching for the target object in the particular region at the second resolution comprises:
determining whether to search for a second anchor object in the particular region, the second anchor object relating to the target object; and
in response to determining to search for the second anchor object is to be detected for in the particular region:
searching, by processing the particular region at the second resolution, for a depiction of the second anchor object in the particular region; and
searching for the target object in the particular region using a result of whether searching the particular region at the second resolution for the depiction of the second anchor object caused detection of a second set of anchor objects in the particular region.
20. A system comprising one or more computers and one or more storage devices on which are stored instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:
accessing image data that comprises an initial image, the initial image depicting a scene to be searched for a target object; and
processing the scene to search for the target object by at least:
detecting, by processing the scene depicted in the initial image, one or more first anchor objects depicted in the scene that are each related to the target object; and
determining a set of regions of the scene by identifying, for each particular anchor object in the one or more first anchor objects, a region of the scene that is relative to the particular anchor object and has a first resolution; and
for a particular region in the set of regions, searching for the target object in the particular region at a second resolution, the second resolution being one of higher than or equal to the first resolution of the particular region at which the particular region was identified.