Building Roof Wireframe

We randomly selected more than 2,100 images with a size of 400×416 pixels from the AIRS dataset as our dataset, And among them, 100 images were used for testing.2,068 images were used to fine-tune the Mask R-CNN network, and 900 of these images were selected to fine-tune the line and junction ‎ extraction model. To create annotation file for building roof wireframe, we drew the real lines of the building roof in Illustrator software and saved them in dxf format.

Then we read this information with the help of the ezdxf library in Python, converted them to needed information and finally wrote them in pickle format. Each pickle file consists of: points (each point is represented by its (x, y) coordinates in the image), lines (each line is represented by the indices of its two end-points), the angle values of branches of each junction, line indexes, the image data, image name and so on.After loading the .pkl file, you can run something like the following in Python to visualize the wireframe: for idx, (i, j) in enumerate(lines, start=0):   x1, y1 = points[i]   x2, y2 = points[j]   cv2.line(im, (int(x1), int(y1)), (int(x2), int(y2)), (0, 255, 0), 2, cv2.LINE_8)I’m excited to share that my work: "Building roof wireframe extraction from aerial images using a three-stream deep neural network" was published on the SPIE Digital Library.The full citation is as follows:Zahra Esmaeily, Mehdi Rezaeian, "Building roof wireframe extraction from aerial images using three-stream deep neural network," J. Electron.