Upload
alagan
View
25
Download
2
Embed Size (px)
DESCRIPTION
Using Artificial Queries to Evaluate Image Retrieval. Nicholas R. Howe Department of Computer Science Cornell University. How Do We Compare Image Retrieval Algorithms?. Different research groups use images from different sources. Image sets are of different sizes. Tasks are different. - PowerPoint PPT Presentation
Citation preview
Using Artificial Queries to Evaluate Image Retrieval
Nicholas R. HoweDepartment of Computer Science
Cornell University
June 12, 2000 Workshop on Content-Based Access of Image and Video Libraries
2
How Do We CompareImage Retrieval Algorithms?
• Different research groups use images from different sources.
• Image sets are of different sizes.• Tasks are different.
– Each researcher identifies set of queries and targets through subjective criteria.
– Can’t share keys because image sets are not standard.
Answer: Badly!
June 12, 2000 Workshop on Content-Based Access of Image and Video Libraries
3
How It’s Usually Done
• Each researcher tests a proposed algorithm against a few baselines.– e.g., Color Histograms.
• No data to compare latest techniques.– Test sets are different.
– Implementation of baselines may differ also.
June 12, 2000 Workshop on Content-Based Access of Image and Video Libraries
4
Some Difficulties
• Given a query, which target is most relevant?
• Context will determine answer.
?
June 12, 2000 Workshop on Content-Based Access of Image and Video Libraries
5
What Should a Good Test Do?
• Provide comparable results even with different image sets.
• Offer insight into the behavior of different retrieval algorithms.
• Run quickly.
• Allow for easy implementation.
June 12, 2000 Workshop on Content-Based Access of Image and Video Libraries
6
Proposal: Altered-Image Queries
f
Image from Library Altered Image
Query
Image Library
1
2
3
etc.
Look for rank of original:
Retrieved ranks:
June 12, 2000 Workshop on Content-Based Access of Image and Video Libraries
7
The Crop Test
• Crop image to k% of its original area.
• Simulates close-up shot of same subject.
Original Crop-50
June 12, 2000 Workshop on Content-Based Access of Image and Video Libraries
8
The Jumble Test
• Shuffle tiles in image divided on an hk grid.
• Simulates image with similar elements in a different arrangement.
Original Jumble-44
June 12, 2000 Workshop on Content-Based Access of Image and Video Libraries
9
The Low-Con Test
• Decrease contrast to k% of its original range.
• Simulates altered lighting conditions and/or camera differences.
Original Low-Con-80
June 12, 2000 Workshop on Content-Based Access of Image and Video Libraries
10
Typical results
• Most retrievals are at low rank.• A few retrievals are at much higher rank.
Median: 26
Mean 205
• Median and mean summarize the results of multiple repetitions.
June 12, 2000 Workshop on Content-Based Access of Image and Video Libraries
11
Difficulty of Altered-Image Queries
• Both mean and median increase with difficulty.• Note order-of-magnitude changes.
June 12, 2000 Workshop on Content-Based Access of Image and Video Libraries
12
How Stable are Altered-Image Queries?
• Ran Crop-50 on three entirely different sets of 6000 images.
• Some consistency even with different test sets.• Look for order-of-magnitude change.
Set 1 Set 2 Set 3 Mean Dev.
Median Rank 5 5 7 5.7 1.2
Mean Rank 29.6 33.9 45.5 36.3 8.2
June 12, 2000 Workshop on Content-Based Access of Image and Video Libraries
13
Does the Number of Images Matter?
• Found linear dependence on number of images.
June 12, 2000 Workshop on Content-Based Access of Image and Video Libraries
14
How Many Queries Must Be Run?
• Small % of total image set gives decent figure.
June 12, 2000 Workshop on Content-Based Access of Image and Video Libraries
15
Comparing Algorithms Using Altered-Image Queries
• Three algorithms compared using altered image queries.
• Especially good or bad performance can be identified.
Crop Jumble Low-Con
Histograms 18 126.6 1 1 86.5 350.3
Correlograms 1 12.4 1 2.0 5 83.6
STAIRS (Tuned) 1 17.0 1 1.2 1 22.6
June 12, 2000 Workshop on Content-Based Access of Image and Video Libraries
16
Final Thoughts
• Altered-Image Queries are...– Well defined.
– Easy to implement.
– Consistent over different image sets.
• A useful addition to our evaluation toolkit.• Also offer diagnostic potential.