Shuihua Wang and Yingli Tian
{swang15, ytian}@ccny.cuny.edu
Presented by: Shizhi Chen
Department of Electrical Engineering
The City College of New York
OutlineMotivation
Proposed algorithm
Experimental results
Motivation Access unfamiliar environment
Recognize restroom signage
Available technology
(a) (b) (c)
Proposed Algorithm
Image Preprocess
Original
Image
Gray
Image
Binary
Image
Connected
Components
Signage Detection: Head Based on shape of Connected Components (CC)
Head shape is circle
Signage Detection: Body More variations
Close to head
Based on shape
Signage Detection Results Scale invariant
Rotation invariant
Illumination invariant
Signage Recognition: Find Corners Search within detected signage region
SIFT (Scale Invariant Feature Transform) detector
Search over all scales
Template
Signage
Detected
Signage
Signage Recognition: Match Corners SIFT descriptor
Histogram of gradients
Rotation and scale invariant
Matching pair of corners: minimal Euclidean distance
Find the template with maximum matching pairs
Template
Signage
Detected
Signage
Signage Database 102 Signage: Men (50); Women(42); Disabled(10)
Experiment Results 89.2% detection rate (91 out of 102 images)
84.3% recognition rate (86 out of 102 images)
Confusion matrix: column is the ground truth
Original
Image
Binary
Image
Connected
Component
Signage
Recognition
W MDW M D
Intermediate Results
MM
MM W
W
WW
W
W
W
W
M
MD
D DD
Recognition Success
Detection Fails
Significant view angle changes
Complex Background
Acknowledgement Supported:
NIH 1R21EY020990,
NSF grants IIS-0957016
EFRI-1137172.
16
Thank you!
Author Contact
Shuihua Wang and Yingli Tian{swang15, ytian}@ccny.cuny.edu
Top Related