I have a document digitalization task where I want to detect technical drawings from images. These Images mostly consist of objects made up of combination of shapes like lines, circles and rectangles. See this example: Example Drawing.

What I want as the result is a detection of all "objects" in this image like e.g. switches, wirings and devices that are present.

I already tried multiple network architectures like:

  • YOLO
  • DeepLab
  • UNet (for Pixel based classification)

Generally I observe that all these approaches work well for small objects but have big problems with "bigger" objects due to their "sparsity". I guess thats not surprising if you consider the nature of CNNs.

One one hand my task seems simple due to things like

  • high contrast
  • limited set of shapes (or "poses") for the different objects

But I think CNNs all have problems with the "sparseness" of the objects as its only the boundary that is detectable and the inner is often empty.

Has anyone here an Idea which architectures to try or links to papers to read for these kinds of problems? Ideally I would like to get the list of objects as output but I am unsure how to encode this in the NN as e.g. YOLO does this by an approach that would not work in my scenario, I think (merging all "inner" boxes that show the same object).

Thanks already!


1 Answer 1


Have you looked at Deformable DETR? These Transformer models are extremely good at detecting big objects because of the Global Attention mechanism.

  • $\begingroup$ Thanks, I will have a look later to see if this could improve my results $\endgroup$
    – Julian
    Feb 16, 2023 at 7:27

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .