Hierarchical Building Recognition

This scheme integrates geometric and appearance information, composed of two stages.

1.  Localized color indexing  More detail

Using two color histograms to represent one image.

Both efficient and discriminative, also tolerate moderate background clutter.

Histogram comparison based on statistics.

Select variable number of candidate based on ambiguity of indexing result.

Based on our experiments with ZUBUD database (201 buildings, 5 training images for each building and 115 test images), 90% recognition rate is achieved in this stage, and 96% have correct model listed in the top 5 candidates list.

Example of selected candidates:

2.  Local appearance based decision

Notice that in the last example, correct model is not the first candidate. We apply a second local appearance based recognition stage to identify true model from candidates list selected by color indexing.

Voting based on keypoints matches already works very well, since first stage discards most candidates which may cause problem for appearance based technique. After applying the second stage, 96% recognition rate is achieved.

The following example shows that the second stage helps identifie correct model.

Our experiments show that the two stages are not only working in a coarse to fine fashion, both also complementary to each other. Let alone the efficiency issue, recognizing with only second stage results in lower recognition rate

Reference:

Wei Zhang and Jana Kosecka, "Localization Based on Building Recognition", 1st IEEE Workshop on Computer Vision Applications for the Visually Impaired, Satellite Workshop of CVPR 2005