Why we need to detect which vertex is convex or concave?

[mridvankeles]

First of all I couldnt fully understand how we are seperating vertecies by looking at their convex_hull? how does that works?

and why we need to seperate them? are we gonna threat convex vertices and concave vertices differently at the inference phase?

Any idea from those situations?

[SarahwXU]

Calculating the convex_hull is implemented as a way to recognize the convex vertices from all of a building's vertices. The convex and concave vertices will be considered as two different classes during the junction heatmap prediction and are processed separately as in https://github.com/SarahwXU/HiSup/blob/45db55433c9ef713dc52b7dd7f7d56e6645202eb/hisup/utils/polygon.py#L26 The separation reduces the ambiguity between nearby building vertices. We did the ablation study about this division of different types of vertices and the results showed improved performance. However, if the pixel coordinates of vertices of building polygons are far apart from each other and easy to identify, it may not need to set two classes.