I know that HOG is pretty state of the art for person detection, but in its original form HOG is only a detection method, and I don't believe it is particularly fast either (there is a significant speedup which takes the sliding window and rejects unlikely possibilities early, but it has been patented).

What is the current state of the art approach for determining the pose of a person (including close poses like a head shot, and hand shot, etc...), and for segmenting the person from the background?

The setting here is still images.

  • $\begingroup$ State of the art requires a depth camera. Think kinect :) $\endgroup$
    – nav
    Dec 10 '12 at 6:13
  • $\begingroup$ Unfortunately the context is that is has to work on existing flat image files. $\endgroup$ Dec 10 '12 at 15:36

Check out Yang and Ramanan's work with flexible mixture of parts. It starts with a central part, but pairs of parts that are attached are detected jointly by estimating the probability that the image area being queried supports the co-occurence of the attached parts in the image. Error is presumed in the attachment and is minimized using a spring-like model. Each part is detected using HOG-style techniques, and all pairs of parts in the human 'tree' are iterated over to determine the detection. In my experience it has a very good false alarm rate, but requires hardware acceleration for real-time applications.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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