Harnessing crowds of humans to train machines

SpaceNet-320x320If you’ve recently searched your phone’s photo library using face recognition technology or read about the recent road tests of self-driving cars, you’ve seen applications of deep learning in action. Most of us still associate artificial intelligence with popular science fiction movies, but many real world applications of for these technologies exist today that are changing how we work, play and live. Deep learning is a form of artificial intelligence that teaches machines to perform tasks through learned experience instead of by programming predefining logic.

This approach mimics the way we as humans learn and is leading to incredible breakthroughs in algorithm performance across a number of fields because of advancements in accelerated computing, open source software, and broadly accessible training data. A terrific example of how this combination of enablers can lead to tremendous innovation is an initiative named ImageNet. Dr. Fei-Fei Li from Stanford University gave a TedTalk in March 2015 on “How we are teaching computers to understand pictures.” It describes how she led a massive effort to create a collection of 15 million publicly available images sorted into over 20,000 categories to advance the field of computer vision technology. Since 2010, an annual ImageNet competition has been held to see who can create the best object detection algorithms.  In 2015 the winning teams built deep learning algorithms that delivered accuracy that for the first time exceeded the performance of most humans.

In the geospatial industry, the commercialization of GEOINT has led to an explosive amount of data being collected to capture our changing planet. Imagine if we could apply deep learning to remote sensing data at scale. DigitalGlobe is working with partners including In-Q-Tel CosmiQ Works and NVIDIA to organize an image mining challenge to engage developers and data scientists to automate the extraction of map features and indicators of activity from satellite imagery.

Today, vector data such as infrastructure and points of interest are primarily mapped through manual or semi-automated techniques. The ability to extract these features accurately and at scale is key to establishing and maintaining current maps. Solving this challenge may enable more advanced use cases, such as change detection, wide area search, automated tipping and downstream uses like autonomous vehicle navigation.

Our long-term vision is to create an equivalent of ImageNet for satellite imagery to advance the application of computer vision in the geospatial domain. To do this, DigitalGlobe and CosmiQ Works have established SpaceNet. SpaceNet will be a collection of satellite imagery and co-registered labeled vector data, which will be made publically available by mid-June 2016. Our goal is to create a sustained collaboration between academia, government and industry to advance the application of artificial intelligence against remote sensing data. In addition to an open resource release, SpaceNet will be used to facilitate machine learning challenges to help advance automated feature identification and extraction algorithms.

Sign up to receive more information about the inaugural SpaceNet Satellite Imagery Object Detection Challenge here: platform.digitalglobe.com/spacenet. Check back in the coming weeks for more information and follow @DigitalGlobe on Twitter.


Challenge sponsorship