Closing Time: CosmiQ Works is Closing Down and Ending its Leadership of SpaceNet
After a run of six years, CosmiQ Works (and The DownLinQ here on Medium) is shutting down. CosmiQ Works’ parent organization (IQT) is folding CosmiQ Works into the IQT Labs organization. Furthermore, CosmiQ Works will no longer act as managing partner and challenge manager for SpaceNet, or be involved in SpaceNet in the future. With CosmiQ Works’ exit, leadership of SpaceNet will be transitioned to co-founder Maxar.
This blog serves as a chance to look back at some of the accomplishments of the CosmiQ Works team over the years (the most recent of which is the acceptance of our SpaceNet 7 paper to CVPR 2021).
1. The CosmiQ Model
Through a series of applied research projects targeted to address national security concerns, CosmiQ Works chipped away at the strategy defined below, providing insight across the workflow chain.
2. CosmiQ Personnel
CosmiQ has never been a large team, varying in size from two to five people over the years. Yet this small team accomplished an impressive amount. Below is a list of the personnel in reverse chronological order of departure, along with an (incomplete) sample of key accomplishments during their CosmiQ tenure.
Adam Van Etten (November 2015 — March 2021)
- YOLT/SIMRDWN, APLS, SpaceNet 3, 5, 7 Challenge Manager
- Now Chief Data Scientist at IQT Labs
Daniel Hogan (February 2019 — March 2021)
- Robustness Study, Introduction to SAR, SCOT (links)
- Now Data Scientist at IQT Labs
Jake Shermeyer (September 2017 — November 2020)
- RarePlanes, SpaceNet 6 Challenge Manager, Super Resolution Trade Study, CometTS
- Now Data Scientist at Capella Space
Ryan Lewis (co-founder, March 2015 — September 2020)
- Head of CosmiQ (2015, 2017–2020), Market Surveys, GM of SpaceNet
- Now at AWS
Nick Weir (September 2018 — April 2020)
- Solaris, SpaceNet 4 Challenge Manager
- Now at AWS
Dave Lindenbaum (co-founder, March 2015 — December 2018)
- SpaceNet Data Manager (Challenges 1–4), SpaceNet 2 Challenge Manager, Commercial SAR Market Survey, Commercial Space Framework
- Now at Accenture
Lisa Porter (October 2015 — October 2018)
- Establishment of SpaceNet, SpaceNet Challenges 1–3, Commercial SAR Market study, Remote Sensing Market study, and the Satellite Communications Market study
- Subsequently Deputy Undersecretary of Defense for Research and Engineering, now co-Founder and co-President of LogiQ, Inc.
Todd Stavish (co-founder, March 2015 —October 2017 )
- Establishment of SpaceNet, SpaceNet 1 Challenge Manager
- Subsequently head of IQT’s Lab 41, now at SRI Ventures
Lee Cohn (February 2017 — October 2017)
- Multispectralization
- Now at Nike
Patrick Hagerty (July 2015 — July 2017)
- SpaceNet Building Metric, Super-Resolution
- Now at Arena
Key Contributing Personnel
Christyn Zehnder
- CosmiQ / SpaceNet marketing, social media, and websites (spacenet.ai, cosmiqworks.org)
Jeremy Joseph
- Problem solving around legal and contract details to keep the SpaceNet collaboration nimble and mission-focused
3. CosmiQ Academic Impact
As an applied research lab, CosmiQ’s primary goal is to demonstrate the art of the possible in government-relevant problems; publishing papers is of secondary importance. Nevertheless, CosmiQ has managed high academic traction for its research projects, with 9 peer-reviewed publications in some of the most prestigious venues (e.g. CVPR, NeurIPS, ICCV) just in the last two years:
- WACV 2019: Satellite Imagery Multiscale Rapid Detection with Windowed Networks (SIMRDWN)
- CVPR EarthVision 2019: The Effects of Super-Resolution on Object Detection Performance in Satellite Imagery
- ICCV 2019: SpaceNet MVOI: a Multi-View Overhead Imagery Dataset
- WACV 2020, City-Scale Road Extraction from Satellite Imagery v2: Road Speeds and Travel Times
- CVPR EarthVision 2020: SpaceNet 6: Multi-Sensor All Weather Mapping Dataset
- IGARSS 2020: Road Network and Travel Time Extraction from Multiple Look Angles with SpaceNet Data
- NeurIPS 2020: SpaceNet 7: Multi-Temporal Urban Development Challenge
- WACV 2021: RarePlanes: Synthetic Data Takes Flight
- CVPR 2021: The Multi-Temporal Urban Development SpaceNet Dataset
4. DownlinQ Stats
Over the past 4.5 years The DownLinQ has gained significant traction as it pushed out blogs at a rapid clip, averaging one blog every fortnight. The most popular blog was published in November of 2016 and has 44,000 views: You Only Look Twice — Multi-Scale Object Detection in Satellite Imagery With Convolutional Neural Networks (Part I). Some overall stats:
See github.com/CosmiQ/DownLinQ_Blog for a full list of DownLinQ blogs.
5. SpaceNet
SpaceNet is an effort dedicated to accelerating open source, artificial intelligence applied research for geospatial applications, specifically foundational mapping (i.e., building footprint & road network detection). SpaceNet was founded in 2016 by CosmiQ Works with Maxar Technologies when both parties recognized that a lack of access to high quality training data was significantly hampering advancements in geospatial machine learning. Since then, SpaceNet has matured into a collaboration with seven additional partners: Amazon Web Services (AWS), Intel AI (a partner 2018–2020), Capella Space, Topcoder, IEEE GRSS, the National Geospatial-Intelligence Agency and Planet.
CosmiQ Works ran SpaceNet and managed seven unique data science challenges focused on foundational mapping (building footprints and road networks). Among other applications, these challenges inform a number of humanitarian and national security use cases where maps may be outdated or non-existent.
In March 2021 as CosmiQ Works gets folded into IQT labs, it will end its leadership and involvement in SpaceNet. With CosmiQ Works’ exit, leadership of SpaceNet will be transitioned to co-founder Maxar. SpaceNet resources (data, code, website) will continue to be openly available to all.
Datasets form the foundation of SpaceNet, specifically high quality imagery and high fidelity labels, as shown in Figure 3.
In order to drive use of these datasets and advance the state of the art, the SpaceNet partners ran a series of challenges, illustrated below.
A few SpaceNet wins:
- Demonstrating that open-source challenges provide value for the US Government. For example, the advancements in road extraction and optimized routing resulting from SpaceNets 3 and 5 connect immediately to numerous national security missions.
- SpaceNet has engaged increasingly talented researchers (who typically lack a geospatial background) to solve difficult geospatial problems. For example, one of the perennial SpaceNet winners (selim_sef) won the massive $1M Facebook Deepfake Detection Challenge.
- Through seven unique challenges, proving that open source competitions can successfully leverage cutting edge machine learning techniques for complex geospatial tasks.
- Providing a massive, high-quality labeled dataset with a permissive license to promote innovation by academia, industry, and the intelligence community.
- Bringing together market players (Maxar, Planet, Capella) as well as academia (IEEE) and government (NGA) into a collaboration to advance geospatial machine learning.
- 8 academic papers by the SpaceNet partners, and 100+ citations to SpaceNet.
- Very high traffic to the dataset (see Figure 5).
5. Conclusion
While there is plenty more we could cover given all the many things the team has done, in the interests of brevity we’ll cut it short here and bid everyone farewell. (Would you like to know more? See our celebratory video and final podcast).
Over the last five years CosmiQ and the DownLinQ have striven be a useful destination for geospatial analytics and research. But now, as the song says: you don’t have to go home, but you can’t. stay. here.
