You are here

Urbanscape: Single Shot Multi-Task 3D Reconstruction

Award Information
Agency: Department of Defense
Branch: Defense Threat Reduction Agency
Contract: HDTRA120C0036
Agency Tracking Number: T2-0403
Amount: $1,098,411.21
Phase: Phase II
Program: STTR
Solicitation Topic Code: DTRA18B-001
Solicitation Number: 18.B
Timeline
Solicitation Year: 2018
Award Year: 2020
Award Start Date (Proposal Award Date): 2020-08-19
Award End Date (Contract End Date): 2022-08-18
Small Business Information
8280 Willow Oaks Corporate Drive Suite 200
Fairfax, VA 22031
United States
DUNS: 078504477
HUBZone Owned: No
Woman Owned: No
Socially and Economically Disadvantaged: No
Principal Investigator
 Yanlin Guo
 Principal Research Scientist/ Principal Investigator
 (609) 356-2386
 yguo@dzynetech.com
Business Contact
 Sarah Wilt
Phone: (703) 291-6661
Email: swilt@dzynetech.com
Research Institution
 Trustees of Indiana University
 Mr. James Becker Mr. James Becker
 
509 E 3rd Street Office of Research Administration
Bloomington, IN 47401
United States

 (812) 855-4884
 Nonprofit College or University
Abstract

Hazard assessment tools that model the transport and dispersion of Chemical, Biological, Radiological, Nuclear and Explosive (CBRNE) materials through urban areas are only as good as the 3D models that inform the physics model. Maintaining accurate, up-to-date 3D models of urban areas is challenging. Even in the commercial world, urban construction and demolition may result in the models created are often out of date for the simple reason that those models are created with LIDAR point clouds or point clouds derived from multi-view aerial or satellite imagery. These data sets are costly to curate and maintain. Recent advances in the area of Deep Learning based 3D reconstruction may be adapted and extended to achieve Urbanscape 3D models from single view satellite imagery. DZYNE’s Urbanscape system will employ deep neural networks to achieve pixel level depth estimation capabilities with semantic and geometric association based on a multi-task learning pipeline that exceed the results of any single task. 3D data annotation scarcity is overcome by training CNNs with a combination of 3D models, DSM, DEM, and multi-view satellite images. A depth map enhanced with semantic labels and geometric properties is created in the format that can be readily incorporated with CBRN tools.

* Information listed above is at the time of submission. *

US Flag An Official Website of the United States Government