indoor/outdoor classification december 1, 2009 liu, cheng yang, hsiu -han han, seung yeob
DESCRIPTION
Indoor/Outdoor Classification December 1, 2009 Liu, Cheng Yang, Hsiu -Han Han, Seung Yeob. Human’s brain is an excellent photo analysis tool and is good at handling high-level information, such as facial recognition. High-level information. - PowerPoint PPT PresentationTRANSCRIPT
12/1/2009 Liu, Cheng Yang, Hsiu-Han Han, Seung Yeob
1
CS 590 Project
Indoor/Outdoor Classification
December 1, 2009
Liu, Cheng
Yang, Hsiu-Han
Han, Seung Yeob
12/1/2009 Liu, Cheng Yang, Hsiu-Han Han, Seung Yeob
2
CS 590 Project
Motivation:Human’s brain is an excellent photo analysis tool and is good at handling high-level information, such as facial recognition.
High-level information
A lot of high-level information to do the indoor/ outdoor classification: Brightness, Background, People, Ground, etc.
12/1/2009 Liu, Cheng Yang, Hsiu-Han Han, Seung Yeob
3
CS 590 Project
Motivation (Cont.):However, human’s brain is not quite convenient for mass data analysis.
will get tired after long run very expensive
In the meanwhile, computers are still indispensable for the analysis of mass data. But not quite efficient for the high-level information.
When processing mass data, can we utilize the low-level information for classification?
12/1/2009 Liu, Cheng Yang, Hsiu-Han Han, Seung Yeob
4
CS 590 Project
Method:
853 1280
853 1280
853 1280
77 7
:
23 28
85 10
:
38 41
87 1
:
15 13
R
G
B
mean_R, std_R
mean_G, std_G
mean_B, std_B
119 indoor photos and 102 outdoor photos were collected.
12/1/2009 Liu, Cheng Yang, Hsiu-Han Han, Seung Yeob
5
CS 590 Project
Low-level Information:Can you tell the differences between these two photos?
The ‘average photo’ for indoor photos
The ‘average photo’ for outdoor photos
12/1/2009 Liu, Cheng Yang, Hsiu-Han Han, Seung Yeob
6
CS 590 Project
Low-level Information (Cont):for Indoor photos for Outdoor photos
Big Difference
Some Difference
Almost the Same
12/1/2009 Liu, Cheng Yang, Hsiu-Han Han, Seung Yeob
7
CS 590 Project
Sampling Method:It is no good to use the whole color matrix to compute the means and stds.
1. low efficiency2. Photos have different sizes. So the sizes of the color matrix
would be different.
12/1/2009 Liu, Cheng Yang, Hsiu-Han Han, Seung Yeob
8
CS 590 Project
Sampling Method (Cont):Uniformly sampling: sample size N = 10,000
12/1/2009 Liu, Cheng Yang, Hsiu-Han Han, Seung Yeob
9
CS 590 Project
A natural question: is the information indeed uniformly distributed on the photos?
Sampling Method (Cont):
12/1/2009 Liu, Cheng Yang, Hsiu-Han Han, Seung Yeob
10
CS 590 Project
Sampling Method (Cont):Non-uniformly sampling: sample size N = 10,000
12/1/2009 Liu, Cheng Yang, Hsiu-Han Han, Seung Yeob
11
CS 590 Project
Classification• Method
– Logistic regression– SVM– Mixture Gaussian
• Samples– Uniform sampling– Non-uniform sampling
12/1/2009 Liu, Cheng Yang, Hsiu-Han Han, Seung Yeob
12
CS 590 Project
Classification(Cont.)
Error rate(%)
Uniform sampling Non-uniform sampling
Logistic regression 17.65% 14.48%
SVM 18.55% 16.29%
Mixture Gaussian 39.37% 37.10%
12/1/2009 Liu, Cheng Yang, Hsiu-Han Han, Seung Yeob
13
CS 590 Project
Classification(Cont.)
12/1/2009 Liu, Cheng Yang, Hsiu-Han Han, Seung Yeob
14
CS 590 Project
Future work
• Improving sampling method– Sample points based on histogram (most frequent values).
• Seeking effective features – We now use purely linear features. Try other feature mapping.
12/1/2009 Liu, Cheng Yang, Hsiu-Han Han, Seung Yeob
15
CS 590 Project
Q&A
Any questions?
Thank you