ONR BAA 08-009

Schedule

• 1 October 08 - Start of project
• 18 July 08 - Notification of funding
• 16 June 08 - Proposal due at ONR
• 13 June 08 (Fri) - FedEx proposal to ONR

Format

• Paper Size – 8.5 x 11 inch paper
• Margins – 1 inch
• Spacing – single or double spaced
• Font – Times New Roman, 12 point
• Number of pages –
  • Volume 1 25 pages. The cover page, table of contents, abstract, executive

summary, and resumes are excluded from the page limitations.

• • Volume 2 has no page limitations.
• Copies – one (1) original, three (3) hard copies, and one PDF copy on CD-ROM

Sections

Volume 1 (see page 15-17 of BAA)

Table of Contents (Not included in page limitations)

Contents of the proposal only, generally by section.

Abstract: (Not included in page limitation)

A brief description of the proposal including goals and objectives, and technology/thrust areas to be addressed.

Executive Summary: (Three (3) page maximum)

A brief summarization of the proposal including the primary areas described below. Emphasis is on the technology in support of FORCEnet, Spiral Development, integration, transition, and relation to other current programs.

Finally, a brief statement why your organization would provide the best value to the government for the particular project.

Statement of Work: (Three (3) pages maximum)

A Statement of Work (SOW) clearly detailing the scope and objectives of the effort and the technical approach. Must include a severable self-standing SOW without any proprietary restrictions, which can be included as an attachment to any resultant contract. When Options are contemplated, the SOW must clearly identify the tasks by separate optional task areas. Similarly, the SOW must include a section listing all the deliverables such as hardware, software, source code, executable code, pseudo code, etc, along with the reporting requirements.

Project Schedule and Milestones: (One (1) page maximum)

A summary of the schedule of events and milestones, with experimentation milestones clearly indicated.

• 1. Port P. Moreel's probabilistic visitation of Lowe's system to industry-grade code.
• 1. Test P. Moreels and Lowe against new database to establish baseline.
• 1. Collect datasets:
  • Airport
  • Sea port
  • Vehicles
  • Buildings along streets of city

1. Annotate datasets for position and identity of objects

• 2. Annotate datasets for viewpoint (front, 3/4, side, 1/4, back, also top and bottom views where applicable)
• 2. Test both systems with varying degrees of resolution
• 2. Test both systems with respect to viewpoint
• 3. (P) Design features that are more viewpoint-invariant.

(P) means `possibly' i.e. probably not

Deliverables: (Two (2) pages maximum)

A detailed description of the results and items to be delivered, including experimentation articles inclusive of the timeframe in which they are to be delivered. Reports and technical items resulting from meetings shall be listed as deliverables (see Section VI, paragraph 2 for required reports and meetings).

Technical Approach: (Ten (10) pages maximum)

Background

1. Viper
2. Caltech work

Challenges

• Quasi-twins (e.g. 2001 Passat vs 2004 Passat)
• The `matrix' problem: each aircraft contains visual info that is airline-specific and other visual info that is model-specific. Can we pull the two apart with supervision? Once the model- and airline- specific characteristics are learnt, can we generalize to new combinations? Can we learn this without supervision?
• Irrelevant differences (e.g. Boeing 747 from Quantas and from United) -- having seen the Quantas, United, American Boeing 747, would we be able to recognize a Lufthansa 747?
• Scaling to 10^6 objects with sublinear costs
  • Efficient representations
  • Efficient searching
• Repeated structure
• Viewpoint change (we will not model 3D -- so this is similar to quasi-twins, how do we avoid having similar features compete). Also need to study how the appearance of features changes with change in viewpoint. Need to generalize from one view to nearby views. Reference P. Moreel's experiments.
• Clutter
• Resolution: down to how many pixels can we recognize things?
• Lighting conditions
• Relatively featureless objects, such as aircraft and boats.
• Making use of known backgrounds (e.g. pictures taken from a stationary surveillance camera) -- how do we integrate Lowe's system with known background.

System Architecture

• The same technique should solve simultaneously various challenges: families of objects, scaling up sublinearly, repeated structure, viewpoint change
• Various methods for scaling up:
  1. Coarse-to-fine
  2. First assign the object coarsely to one or more families of objects (perhaps using visual words), then use David Lowe's scheme restricted to those families. Perhaps need to go down a number of family branches simultaneously as in Fleuret and Geman in their face finder).
  3. Use probabilistic scoring to take soft decisions and keep multiple hypotheses going simultaneously (as in P. Moreels' paper).

Data Collection

1. Use available datasets
   • Recognition Benchmark Images from University of Kentucky here
   • The Oxford Building Dataset here
   • Amsterdam Library of Object Images here
2. Collect from Flickr using appropriate keywords (probably not because of IP problems).
3. Collect pictures of jetliners from JetPhotos
4. Collect three datasets from fixed positions:
   • Boat traffic exiting-entering Marina del Rey
   • Car traffic along fwy (e.g. 210 in Pasadena, traffic data available here in the 2007AADT report)
   • Plane traffic at LAX (locations described here).

In each case collect data many days in order to have sufficient repetition of specific vehicles, boats, planes. Ground truth by hand. Also helped by Airline schedules (see e.g. floweb.com). Unclear how to get ground truth on boats -- need an expert on sailing and motorboats here. Enlist plane spotters (e.g. from Aircraftspotting.net) to help classify the pictures. See also Wikipedia Aircraft Spotting page.

Need to be specific on the n. of distinct vehicles / boats / aircraft we expect to see. Also the total n. of vehicles.

One aspect that seems important is the existence of quasi-twins that have some differences (e.g. 2001 Passat vs 2004 Passat, which may have slightly different headlights and back lights -- this actually may make for a good picture in the proposal).

Past Performance: (Two (2) pages maximum)

Past performance will consist of a description of the offeror’s Government contracts (both prime and major subcontracts (those involving 25% or more of the effort)) received during the past three (3) years), which are similar to the effort being proposed. The offeror may describe any quality awards or certificates that indicate the offeror possesses a high quality process for providing desired research and development outcomes.

Other Agencies: (Not included in page limitation)

CALTECH: NONE

Useful Information

• BAA-2008-09, ONR BAA page