Problem statement

I am starting a project soon where I need to create a method to automatically create landmarks on a set of images of human hands. I do not have a sample image but the images are taken with the same background and stable lightning conditions. The images probably look similar to this one, (but should be the other side of the hand).

sample hand

My colleagues have defined some landmarks, but I would like to first just detect the tips of the fingers and the inner most part between the fingers. I was looking into potential solutions and active shape or active appearance models look like a good idea, although they might also be an overkill. I was also thinking about just creating a template of a hand with landmarks, segmenting the hand with foreground/background segmentation (probably thresholding, PCA or even MRFs if the others fail a lot), and then do an affine + non-rigid registration to transfer the landmarks.

What I am asking is:

  • Are the methods that I mention for solving this task appropriate?
  • Is there any solution out there that already solves this problem?
  • Referrals to code to do some of this is appreciated.
  • Do you have other better suited ideas to solve this problem?


I am going to play around with the images in this dataset, which already have lots of annotated landmarks.

I'm new

Yes, this is my first post, so if you have ideas for other tags to include let me know. Also if you need further clarifications, don't hesitate to ask!


1 Answer 1


There are two simple approaches.

  1. Find the hand region with thresholding and then use something like convexity defects. Even though you will have many spurious landmarks at the end, you could easily cluster them out, as your scene is well controlled. There are also couple of youtube videos, if you like to get an idea in advance, [1], [2].

  2. Fit a hand model to the image. There are many works in this field, and it's an active academic area. Some are, [1], [2], [3]. Once the model is fit, you could extract any landmark you desire, since the correspondence to the canonical hand frame is determined.

The pdf you provided also gives a simple method for fitting a hand region.

  • $\begingroup$ Yeah, I have been going through the pdf. My only concern was if there was a simpler way to do this. Thanks for your answer! I'll look into the references. I'm going to wait and see if more people comment on this. Otherwise I'll accept your answer! $\endgroup$
    – Gumeo
    Apr 2, 2016 at 15:44
  • 1
    $\begingroup$ Updated the answer with videos. $\endgroup$ Apr 2, 2016 at 17:00

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.