VIRAT 2008

From Vision Wiki
Jump to navigation Jump to search

Team 1: Kitware

Participants:

  • Anthony Hoogs (Kitware) - coordinates overall
  • Trevor Darrell (ICSI, UC Berkeley) - Coordinates UCBerkeley (Darrell, Malik) and Caltech
  • Honeywell, General Dynamics, CMU, Maryland, U. Texas ...

Budget: 75K$ per year, 18 + 12 + 12 months

To Do List

Team 2: BAE

Participants:

  • Joel Douglas, Director, Computer Vision Group, BAE
  • Trevor Darrell (ICSI, UC Berkeley) - Coordinates UCBerkeley (Darrell, Malik) and Caltech

Budget: 75K$ per year, 18 + 12 + 12 months

Proposing: Crowd-level activity recognition.

To Do List

Team 3: Northrop Grumman

Participants:

  • Paul Smith, Principal Investigator, Northrop Grumman
  • Dave Newburn, Northrop Grummann - Proposal Manager


To Do List

Utopia Compression

  • Hai Wei, Utopia Compression- Main POC. (he will integrate Caltech's proposal with rest)

hieu@utopiacompression.com Tel: 310-473-1500 x.105)


Budget:??

  • Shooting for 1 post-doc + odds and ends = 130K$

Notes: Apr 22 08 |

General Info

  • Name: Pietro Perona
  • Organization: California Institute of Technology (Caltech)
  • Degree: PhD, 1990, University of California at Berkeley
  • Title: Allen E. Puckett Professor and Professor of Electrical Engineering
  • Other: Executive Officer (Chair), Computation and Neural Systems, Caltech


  • Prior work relevant to Virat (venue of representative publication in brackets):
    • Representation of visual categories (FG05, CVPR96, CVPR98)
    • Learning visual categories with weak or no supervision (ECCV00, CVPR03)
    • One-shot learning of visual categories (ICCV03)
    • Constellation models for human motion classification (ECCV00, PAMI03)
    • Learning movemes without supervision (CVPR05, C. Fanti's PhD thesis, 08)
    • Model of visual saliency (NIPS06)
    • Interplay of generative and discriminative models for recognition (IJCV08)
  • Narrative on prior work: Visual `events' or `objects' (actions, activities, objects, scenes) are naturally grouped by humans into categories. In order for machine to categorize visual events we need to address three issues: (a) how to model categories, (b) how to match models to images (detection, classification), (c) how to learn models from training examples. Professor Perona's research centers on these three issues, with a special focus on the learning problem.
  • Facilities: Professor Perona's laboratory occupies approximately 2000sqft in the Moore Building at Caltech in Pasadena. The laboratory includes office space for the PI, an assistant and 12 post-docs and students, as well as a meeting area, a psychophysics room and a hardware work area. Equipment includes 18 high-end workstations, a computational cluster with 30 CPUs, 4Tbytes disk space, a 50Hz human motion capture system with 6 cameras and IR lighting fixtures.


  • Most related previous projects:
    • 2000-05 - MURI on recognition of actions and activities (subcontract from UC Berkeley)
    • 2005-10 - MURI ``Learning to Recognize for Visual Surveillance (Caltech is prime, other sites: UCLA, MIT, UCBerkeley, UCIrvine, U. Illinois)