Recognition of Physical Changes in a Building
 

We started by obtaining a set of 22 images which were taken using different lighting conditions. The position of the camera was kept constant during the acquisition of the images. As you can see, the training set has a hide variety of lighting conditions.

The input images were normalized using two different mean and standard deviation. The following pictures show the mean image obtained using the values bellow. The picture on the left is less vulnerable to light changes. On the other hand, the one on the right gives a more clear image.

mean = 100     std = 80                                   mean = 100     std = 100

The tests that were carried were performed using the second case (100,100). Bellow is an explanation of the type of changes that were made to the original model and the results obtained. You can click on the image number to see the input image.

Test #	Description of the change made to model	Max distance	Min distance
1	1 small pin was removed (Upper left)	39253	36720
2	2 small pins were removed	40539	37848
3	4 small pins were removed	48114	44846
4	1 white pin, 1 small pin (upper right)	47576	44556
5	2 white pins (upper left and right)	49145	46473
6	Lights off	52783	49967
7	1 blue rod was removed (lower right)	48300	45623
8	1 small pin was added (middle)	47640	44640
9	1 white star was removed (left)	47725	44506
10	1 blue star replaced the white one	47361	44178
11	2 white pins were replaced with 2 gray ones (bottom)	48378	45285
12	1 white star was removed (lower right)	47559	44394
13	4 green small pins were changed by black ones	48321	45077
14	1 orange bar was added (lower right)	47602	44349
15	1 piece was removed from the right edge	47507	44123
16	same as 15 but with different background color	47507	44123
17	upper right corner was substituted by a different one	48037	44746
18	upper right corner was removed	47802	44482
19	3 red pieces representing fire were added (middle right)	47890	44860
20	2 fires in the middle section	48034	44834
21	1 white pin was added in the middle section	52800	50266
22	1 piece was removed from the buttom edge	47865	44633
23	traiming set image # 3	37875	34749
24	traiming set image # 18	40657	34703
25	traiming set image # 21	39296	36349
26	images not in the training set	48346	45430
27	images not in the training set	48168	43624
28	images not in the training set	50923	46194

These images will give you the overall idea of how the tests were performed.

Discussion:
The first 22 images agree with what was expected; as more parts are removed, the Euclidian distance increases. Images 19 and 20 simulate a case when fire occurs. When looking at input images that are within the training set, 2 out of the three images (23-25) are greater than the maximum expected value of  39253. This value was chosen based on the results of the first input image. We should note that they were performed under different lighting conditions. Finally, if the images are not within the training set (26-28) the distance is considerably higher, even though they didn't have any defects. One of the explanations, is that the shadow of the wall is sufficiently different from the images in the training set, yielding a large Euclidian distance.

 

Back