To leverage the commercial sector’s progress in artificial intelligence, the Pentagon’s Defense Innovation Unit Experimental (DIUx) announced a contest awarding a total of $100,000 to creators of effective computer vision tools. The xView Detection Challenge calls for skilled developers worldwide to submit object detection algorithms built for disaster response applications. DIUx will begin accepting entries next month.

The challenge is a collaboration with the National Geospatial-Intelligence Agency (NGA) in support of the Agency’s work providing relief in the wake of disasters such as Hurricanes Harvey and Irma. Reducing minimum image resolution and improving detection of fine-grained object classes are among the challenge’s core focuses.

Registered contestants will train their programs using xView, one of the Pentagon’s largest publicly available datasets of hand-annotated overhead imagery. The dataset covers more than 1,400 square kilometers in both visible and infrared light, and contains roughly one million examples of 60 diverse object classes. Features the algorithms will be required to learn—such as damaged buildings, vehicle lots, utility trucks, tanker trailers, and excavators—are marked by bounding boxes. The specified objects were chosen based on their relevance during humanitarian aid efforts, WIRED reports, and are specific enough to push the recognition capabilities of existing programs. While the training dataset will be released under a public, noncommercial license, the final testing dataset will not.

Winning algorithms—those with the highest accuracy and most complete coverage—will remain freely available for use by DIUx and NGA, who may continue development internally or re-train the software for other purposes such as warfighter support. Winning developers may be called on for follow-up work supporting other national security missions. DIUx has yet to reveal how the total cash prize will be divided among winners, but according to WIRED, a $5,000 special prize will go to the best open-source entry as a way to encourage strings-free sharing. Additional prizes will be awarded in the form of cloud compute credits.

After the challenge ends in May, DIUx will host a workshop featuring representatives from academia, industry, and government, who will discuss challenges and breakthroughs that arose while working with the xView dataset.

xView builds on the foundation laid by other challenges such as IARPA’s Functional Map of the World and DigitalGlobe’s SpaceNet series, which set out to accelerate machine learning capabilities using geo-tagged data. If computer vision programs could assume the task of scanning data and identifying damaged homes or highway impasses, first responders on the ground would receive near real-time situational awareness with greater speed and accuracy. Meanwhile, human analysts could redirect their energy and critical thinking skills toward more demanding tasks.

Photo Credit: DIUx


Posted by Andrew Foerch