Hybrid image templates learned from roughly aligned image examples
In the first experiment, we apply the learning algorithm to automatically learning HiT's from image categories. Below are the hybrid image templates learned from object or scene categories. The number of training images for each category varies. Most categories have around 30 images, and some categories have as few as 6. The selected patches are ranked and selected by their information gains. Sketch and texture patches are selected for most categories. Flatness patches are selected for tomato and the sky areas. Many color patches are selected for tomato, pizza, highway and forest, and some are chosen for panda, palm tree and sunset.
| examples | structure | texture | flatness | color | |
|  |
 |
 |
 |
|
|  |
 |
 |
 |
|
|  |
 |
 |
 |
|
|  |
 |
 |
 |
|
|  |
 |
 |
 |
|
|  |
 |
 |
||
|  |
 |
 |
||
|  |
 |
 |
 |
|
|  |
 |
 |
 |
|
|  |
 |
 |
 |
|
|  |
 |
 |
 |
|
|  |
 |
 |
 |
|
|  |
 |
 |
 |
|
|  |
 |
 |
 |
|
|  |
 |
 |
 |
|
|  |
 |
 |
* For presentation purpose we only display templates of one scale.
The learning method follows directly from the active basis model. It is formulated as a step-wise variable selection and parameter estimation, which we may call information projection.
Structure, texture and flatness features are all defined on a common dictionary of Gabor elements at 16 orientations at different scales.