Transcript Building Recognition

Building Recognition
Landry Huet
Sung Hee Park
DW Wheeler
Problem Statement
• Identify Stanford buildings
from photos
– 16 buildings
– Database of 300 pictures
• Fast enough to implement
real time system
Project Outline
Image
database
Bldg name
color
histogram
1. Color histogram matching
2. SIFT feature matching
3. Image-by-image comparison
color
histogram
List of
SIFT
descriptors
Image descriptor
Feature
database
Ransac
Skilling
Img #
Bldg
SIFT
descriptor
Feature descriptor
Approach and Results
• Timing speed-up
– Find buildings in database that
have similar color properties
– Use kd-tree to find images with
the most SIFT feature matches
– Time reduced from 34 seconds to
22 seconds
Approach and Results
• Accuracy improvement
– Distinguish buildings by both color
information and SIFT features
– Use HSV color representation and
color normalization to be invariant
to light conditions
– Measure average error between
inlier features using ransac
algorithm
Work Distribution
• Landry Huet
– Feature space search, kd-tree structure,
photography
• Sung Hee Park
– Database interface, SIFT matching,
Ransac, vanishing points, photography
• DW Wheeler
– Color histograms, photography