We use four initial placements in our experiments, although two initial placements usually suffices (as shown in the results for the initial placement phase).
For 20 return images, SEDL achieved a recall rate of 78% over the 25 queries. This means that 78% of all advertisements that should have been retrieved over all 25 queries were retrieved in the top 20. Of course, a query result with the three breathe right advertisements returned as 1st, 2nd, and 3rd is more precise a result than if the three are returned 1st, 5th, and 20th, which is more precise than if the three are returned 18th, 19th, and 20th. The precision is a number between 0 and 1 which measures how close to perfect ranks the correctly returned images occurred, where 1 is perfect precision. The precision of SEDL over all 25 queries is 0.88. With 20 return images, SEDL achieved high recall and high precision. The average time for one query is Tavg=40 seconds, which is about 0.1 seconds per query-image comparison.
Some SEDL query results are shown below. The query time and the number of correct advertisements returned in the top twenty are shown below each result. For the first five queries, SEDL achieved perfect recall and precision.
|The three breathe right ads are returned 1st, 2nd, and 3rd.|
|The two comet ads are returned 1st and 2nd.|
|The two taco bell ads are returned 1st and 2nd.|
|The two jello ads are returned 1st and 2nd.|
|The two fresh step ads are returned 1st and 2nd.|
The clorox query result is perfect recall, but not perfect precision.
|The five clorox ads are returned 1st, 3rd, 4th, 7th, and 8th.|
The Apple query result is neither perfect recall nor perfect precision.
|One of the three Apple ads is not returned in the top twenty. The
remaining two Apple ads are|
returned 1st and 4th.
|top||Title, Table of Contents, Introduction|
|prev||The Verification and Refinement Phase|
|next||Experiments with a Shape Database of Chinese Characters|
The ideas and results contained in this document are part of my thesis, which will be published as a Stanford computer science technical report in June 1999.
S. Cohen. Finding Color and Shape Patterns in Images. Thesis Technical Report STAN-CS-TR-99-?. To be published June 1999.
Email comments to firstname.lastname@example.org.