I have been researching about road surface segmentation. Some things that I noticed:
- Most of road segmentation research that has been done applies the segmentation on either front-view perspective images (autonomous cars) or top-down perspective satellite images
- They fail to detect more than one separate roads if there are multiple roads in an image
I am trying to extract every road surface area from a traffic scene. For example traffic scenes like this:
And expecting this kind of result:
I have tried various autonomous vehicle road segmentation model and they always fail to detect multiple roads in the scene. Sometimes, they also fail to detect anything because I suppose the models are trained with front-view perspective scenes only. I have not tried the top-down satellite perspective model, but I reckon it will also give bad result because of perspective difference.
My question is, is there any code reference or research about multiple road extraction from traffic scenes? Or even, is it even possible at all with images from this perspective? The image source refer to a paper that does exactly what I want, but the explanation of their approach still eludes me. I'm not sure if the process involves deep learning or manual segmentation process.
Thank you.