In this final step, we use our opening segmentation results from EQanalytics/Data/Opening_Detection_Results/Opening_Results.csv to determine a softness score. To execute the classification run
python Classifier.py
As a first step, the classifier filters all elements which are not labeled door, window or shop. Then it removes element thats have large overlap and keeps the elements with the bigger area.
Next, the classifier uses K-mean clustering to determine the number of stories. To do so, the algorithm looks at the y-coordinates of all segmented objects and finds an ideal fit of clusters. The results of the level detection will be saved to EQanalytics/Data/Level_Detection_Results. By using the flag --no_save_img the detected level images will not be saved as separate images.
Finally, once all segmented objects are determined, we calculate a quotient of the total width of openings on the second level over the total width of openings on the first level. To avoid double counting only the interval union of all widths are considered. The results, along with their street address, are saved to EQanalytics/Data/Softness_Scores.csv.
| Score x | Class |
|---|---|
| 0.3 < x <= 0.75 | soft |
| 0.75 < x <= 1.5 | non_soft |
| x > 1.5 | x <= 0.3 | undetermined |
If the softness score is too high or too low, then there may be some issues with obstructed views and we err on the side of caution and return undetermined. Otherwise, if it is less than 75%, we classify the building as soft and non_soft otherwise.
This completes our vulnerable building detection model. The Softness_Score.csv file contains the classification for all images that were processed.