Background segmentation

I do not know if it is the right forum, if not, I apologize.

You see I’m trying to find some architecture to segment foreground / background. Initially I thought of a multiclass segmentation model learned with COCO + CRF (DeepLabv3 +). The problem is that, in my opinion, this approach is not good enough. Well, the model has learned a different problem.

The medium in which the algorithm would work works in static camera and I can perfectly obtain a reference image of the background. The question is: Is there an architecture that only separates foreground/background using a background image as a reference? That the algorithm is capable of segmenting objects that it does not know?

Obviously you could use classic background subtraction: MOG2, KNN, etc … The problem is that they are too sensitive. Is there a more robust analog in Deep Learning?

Thank you very much!

1 Like