ORE evaluation
Engine optimized for tests with "random" pictures from Google(considering random background, orientation, scale, position) and gives following results in Select From scenario:
15/02/2010
| Classes | Number of pictures tested | Accuracy |
Airplanes |
1464 |
97.68 |
Faces |
218 |
98.17 |
| Motobikes | 399 | 99.00 |
Engine optimized for cases of "rotation" of image, tested with Caltech-101 image base. Below is comparison of accuracy with original testing set and same set rotated on 15 degrees. Classes Accuaracy with original Accuracy with rotated on 15 degrees Airplanes 98.50 99.75 Faces 98.17 95.41 Motobikes 99.0 85.21
10/02/2010
| Classes | Accuaracy with original | Accuracy with rotated on 15 degrees |
Airplanes |
98.50 |
99.75 |
Faces |
98.17 |
95.41 |
| Motobikes | 99.0 | 85.21 |
Considerably improved results for SELECT ONE scenario. Tested with Caltech-101 image base and compared with publically available sources.
20/12/2009
| Classes | |||
|---|---|---|---|
| Airplane | Motorbike | Face | |
TomskLabs |
99.0 |
99.75 |
100 |
Csurka Gabriella |
96.3 |
92.7 |
94 |
Josef Sivic2 |
97.5 |
96.5 |
99.54 |
- Numbers in cells represent percentage of correctly detected images from the testing set (accuracy)
- Other published results are given for comparison:
- http://www.cs.cmu.edu/~efros/courses/LBMV07/Papers/csurka-eccv-04.pdf
- http://www.di.ens.fr/~russell/papers/Sivic05.pdf