I have a set of polygons for each image. Those polygons consist of four $x$ and $y$ coordinates. For each image, I need to extract the ones of interest. This could be formulated as an Image Segmentation task where, for example I want to extract the objects of interest, here: cars.

enter image description here

But since I already get the polygons through a different part of my pipeline I would like to create a simpler machine learning model. The input will not be the image but only the coordinates of the polygons. In this model each sample should consist of multiple polygons (those can vary in number) and the model should output the ones of interest.

In my mind, I formulated the problem as follows:

  1. The polygons are the features. Problem: Samples will have varying number of features.
  2. The output will consist of the indices of the "features" (polygons) I am interested in.

First, I created a decision tree and classified each coordinate as $0$ (not interested in) or $1$ (of interest). But, by doing this, I don't consider the other coordinates that belong to the image. The information of the surrounding is lost.

Does someone have an idea of how to model this problem without using Image Segmentation?

  • 1
    $\begingroup$ Can you please explain in more detail, how the coordinates are selected? what makes them interesting and how is the relation between the coordinates and the sample? Does the "interestingness" of a coordinate in any way depend on its assignment to any sample? Supervised learning generally means, you can train your model with a set of data for which the decision is known. So you have such a labelled data set, right? $\endgroup$
    – jottbe
    Jul 2, 2019 at 23:53


You must log in to answer this question.

Browse other questions tagged .