Visualisation of Inference

[Ankit-Vohra]

Hello Author,
First of all want to comment your approach looks promising and I want to test it
I'm facing a couple of issues

1. How to store the inference visualisation?
2. Can we convert the output polygon into ESRI Shapefile or GeoJSON?
3. Is it possible to store the black and white mask?

[SarahwXU]

Hi,
Thanks for your interest.
For the question 1, you can look into the function “save_viz” from utils/visualizer.py.

HiSup/hisup/utils/visualizer.py

Line 44 in c1a201b

def save_viz(image, polys, save_path, filename):

For the question 2, it is sure that the output polygons could be converted into ESRI Shapefile or GeoJSON as long as you have the geographic coordinates. But the public AICrowd dataset does not provide that, so we do not give the code.
For the question 3, the predicted mask is acquired here:

HiSup/hisup/detector.py

Line 121 in c1a201b

remask_pred = remask_pred.softmax(1)[:, 1:]

. You may set a breakpoint or add the save code.

[Ankit-Vohra]

Is there any way I can use my personal developed model and just try the image post processing work which you are doing to regularise the building edges

[SarahwXU]

First of all, I am not sure that I fully understand your meaning. A detailed description about the function of your personal developed model would help. Our work of HiSup emphasizes the interaction among the three representations (vertices, line segments, masks) and the joint training under all supervisions boosts the performance. From what I understand, you could add the supervisions from HiSup to your model, therefore, get a regular polygonal mapping.