webdrag0n / building-feature-recognition

Using computer vision method to recognize and calcutate the features of the architecture.

building-feature-recognition

In this repository, we accomplished building feature recognition using traditional/dl-assisted computer vision method. The Chinese version of the README is here. And the report of our project(in Chinese) is in the report folder.

Results

number of floors

	Business school	dorm5	network building
ours	5	6	6
net	5	6	6

number of windows

	Business school	dorm5	network building
ours	60	87	44
net	60	100	44

size of windows

	Business school	dorm5	network building
ours	L6.0m-W1.9m	L3.2m-W1.9m	L2.6m-W1.9m
net	L4.7m-W2.1m	L3.4m-W1.7m	L2.6m-W2.3m

area of windows($m^2$)

	Business school	dorm5	network building
ours	12	6	5
net	9.9	5.9	5.9

max length($m$)

	Business school	dorm5	network building
ours	120	35	40
net	126	36.7	45.7

max width($m$)

	Business school	dorm5	network building
ours	60	15	18
net	69.7	16.3	17.9

max height($m$)

	Business school	dorm5	network building
ours	31	22	19
net	26.2	21.1	24

Floor area($m^2$)

	Business school	dorm5	network building
ours	7600	1800	700
net	7630	1599	743

volume($m^3$)

	Business school	dorm5	network building
ours	23.6	3.96	1.47
net	19.9	3.37	1.78

Using the code

To use our code, please first clone this repository and install the cv2, numpy, matplotlib package.

Reproduction

side_main.py process pictures taken horizontally and count floor numbers as well as their relative proportion.

top_main.py process pictures taken by a drone from the top of the building.

window_main.py process pictures taken horizontally and count window numbers as well as their relative proportion.

Simply click the 'run' icon when opening the three files and you will see the results. Note that threshold tuning might be needed, please refer to our report for more details.

The demo pictures and the results are in the demo folder.

Test your own picture

Please note that the computer vision methods are sensitive to the quality of the picture.

To test your own picture, you may simply replace the picture name in the code. If the default thresholds don't work well, just replace it with a few tests.