Visual recognition software has developed to the point where the Defense Advanced Research Project Agency believes the private sector could deliver a tool capable of extracting a range of information from fuzzy, noisy and ill-posed images that have little or no metadata and hail from unknown sources in an unknown context.

“Our adversaries frequently use video, still and cell phone cameras to document their training and operations and occasionally post this content to widely available websites,” the agency says in an Aug. 26 solicitation for its Visual Media Reasoning program.

“The volume of this visual media is growing rapidly and is quickly outpacing our ability to review, let alone analyze, the contents of every image,” it adds.

The tool DARPA is after would be able to extract the “who, what, where and when” from within an image. The interplay of possibilities between those elements could help inform each other, the solicitation says. For example, a certain time of year identified in a picture could help confirm the identity of a recently-released insurgent in the same photo.

The solicitation acknowledges that no 100 percent automated tool would likely satisfy user needs. Human interaction with the tool might require the user to verify a rough approximation of the distance “between a mountain and a building, or to draw-in their best guess as to the shape of a partially occluded building roofline.”

Visual recognition software has developed to the point where the Defense Advanced Research Project Agency believes the private sector could deliver a tool capable of extracting a range of information from fuzzy, noisy and ill-posed images that have little or no metadata and hail from unknown sources in an unknown context.

“Our adversaries frequently use video, still and cell phone cameras to document their training and operations and occasionally post this content to widely available websites,” the agency says in an Aug. 26 solicitation for its Visual Media Reasoning program.

“The volume of this visual media is growing rapidly and is quickly outpacing our ability to review, let alone analyze, the contents of every image,” it adds.

The tool DARPA is after would be able to extract the “who, what, where and when” from within an image. The interplay of possibilities between those elements could help inform each other, the solicitation says. For example, a certain time of year identified in a picture could help confirm the identity of a recently-released insurgent in the same photo.

The solicitation acknowledges that no 100 percent automated tool would likely satisfy user needs. Human interaction with the tool might require the user to verify a rough approximation of the distance “between a mountain and a building, or to draw-in their best guess as to the shape of a partially occluded building roofline.”