Featured Computer Vision Programs

For more than ten years, Kitware has been involved in major, national R&D programs funded by the Defense Advanced Research Projects Agency’s (DARPA) and Intelligence Advanced Research Projects Activity (IARPA). On these highly competitive, high-profile programs, Kitware has developed state-of-the-algorithms, integrated them into prototype systems, supported third-party evaluations, and coordinated large teams with multiple universities and industry subcontractors. A selection of our prominent R&D programs are described here.

MediFor

Media Forensics

DARPA’s Media Forensics program aims to develop a platform that can automatically detect video and image manipulations, provide detailed information about how these manipulations were performed, and investigate the overall integrity of visual media. Kitware developed novel capabilities to detect all types of image manipulations in a unified approach, and frame-level video manipulations such as deleting or duplicating sequences of frames. Kitware led a world-class team that included Columbia University, Dartmouth College, SUNY Albany, University of California Berkeley, and the University of Oregon. This project started in 2016 and will be completed in 2021.

VIAME

Video and Image Analytics for the Marine Environment

VIAME is an open source computer vision software platform designed for do-it-yourself artificial intelligence (AI) for analyzing imagery and video. Originally developed for analytics in the maritime domain, it now contains many generic algorithms and capabilities that apply to virtually any video or image domain. VIAME enables users to utilize and customize deep learning algorithms for their specific analytic problems, without any knowledge of programming or AI. VIAME is funded by the Automated Imagery Analysis Strategic Initiative within the National Oceanic and Atmospheric Administration (NOAA) National Marine Fisheries Service. This project started in 2015 and is still ongoing.

DIVA

Deep Intermodal Video Analytics

Funded by IARPA, the DIVA program is developing automatic activity detection for a multi-camera streaming video scenes. Performer teams are developing systems to detect 37 known activities, including person actions, person-vehicle interactions, in indoor and outdoor locations. As the test and evaluation support contractor, Kitware collected a very large dataset consisting of dozens of video cameras, UAV video, GPS tracks and more than one hundred participants in realistic settings. Kitware also annotated hundreds of video hours with space-time bounding boxes of dozens of activity types. This project started in 2016 and is ongoing.

XAI

Explainable AI

The XAI program aims to create a suite of machine learning techniques that will produce more explainable models while maintaining a high level of accuracy and enable human users to understand, appropriately trust, and manage their AI partners. Kitware is a subcontractor to the University of California, Berkeley for this project funded by DARPA. Kitware has developed methods to help explain deep learning methods for content-based retrieval and reverse image search, culminating in a prototype and user studies showing that visual explanations improve user accuracy and trust in the AI system. This project started in 2017 and is still ongoing.

CORE3D

Computation of Operationally-Realistic 3D Datasets

Funded by the Intelligence Advanced Research Projects Activity (IARPA), the CORE3D program developed technology that automatically generates accurate 3D object models using satellite imagery. Kitware’s effort focused on converting 3D point clouds to low-complexity 3D surface meshes, and recognizing surface material types. Kitware led a world-class team that included Columbia University, Purdue University, Raytheon Technologies, and Rutgers University. This project started in 2018 and was completed in 2020.

VIRAT

Video and Image Retrieval and Analysis Toolkit

Funded by DARPA, the VIRAT program developed capabilities for automatic detection and real-time alerts of events and human actions in aerial surveillance video, both visible and IR. The system also indexed the descriptors into a database to enable subsequent search for similar and related events. Kitware led the effort after the phase 1 downselection and supported subsequent integration and extensive testing on operational data. Kitware led a world-class team that included California Institute of Technology, Columbia University, Cornell University, General Dynamics, Honeywell, Rensselaer Polytechnic Institute, University of California Berkeley, University of Maryland, and the University of Texas Austin. In later phases, Kitware was a subcontractor to Lockheed Martin on transition efforts. This project started in 2008 and was completed in 2012.