Project Goal
Will use brightness normalized Scale Invariant Feature Transform (SIFT*) as features fed into a Fergus, Perona, Zisserman** style learned model of people (trained from background subtracted images outdoors). Will then find and track people from hand held camera (or recorded Video Cam tape) while moving outdoors.
Figure 1: Unsupervised learning of face models.**
Project Scope
Will use the existing Matlab SIFT code developed at Intel as a basis for feature finding, recognition keys will have to be added to the features found by this code. SIFT features might also have to be modified in method suggested in Project 9 for illumination invariance. Fergus et al**'s method will be used, except with SIFT features. Alternatively, a graphical model's expert can propose a new method of embedding SIFT observations in a graphical model for tracking people. Data of people outdoors will be taken for training using a Video Cam. Further tests sets live or from a Video Cam will demonstrate identification of people outdoors.
Tasks
Project will be considered complete at Task 5. Bonus for further work..
- Task 1: Download, run and familiarize yourself with the Matlab SIFT (0.3M) code.
- Task 2: Read background papers in SIFT original (0.5M) and current (0.5M). Also read Fergus et al (3.6M) and background Weber (0.8M)
- Task 3: Collect walking VideoCam data of people outdoors at several different times of day/illumination.
- Task 4: Train up recognition of people under many different views and backgrounds. Might need to separate out widely different views into frontal, side, back etc. A "pose" parameter could be introduced if method can be done using Probabilistic Graphical Models.
- Task 5: Identify people on new "walk around campus". Track recognition rate, false positives and negatives.
- Task 6: Identify front, back, side.
- Task 7: Identify male/female.
- Task 8: Demonstrate it working indoors.
Pre-requisites
It would be great if at least on of the students were from Koller's group or had taken her Bayesian Net/Graphical Model's course to attempt to put people detection with SIFT features in a graphical model form.
Project Contact
Project Status
three free slots
Midterm Report
not yet submitted
* David G. Lowe, "Object Recognition from Local Scale-Invariant Features",
ICCV'99
** R. Fergus, P. Perona A., Zisserman, "Object
Class Recognition by Unsupervised Scale-Invariant Learning", 2003
