0
$\begingroup$

FindFaces doesn't work when people are far away from the camera. I'm looking for a single accurate method of detecting one or more people in an images, at all scales.

|improve this question
$\endgroup$
  • $\begingroup$ It was me who downvoted and I won't explain myself because you are not a newbie. Anyway: Object detection and localization using neural network, don't know if it works 'accurately at all scales'. $\endgroup$ – Kuba♦ Feb 27 '19 at 13:41
  • $\begingroup$ No worries, I think this is a good question to have asked separately because human detection is super important and pedestrian accuracy is still not great in the old yolo v2 net in the answer that you linked. I guess I will compare using that older netmodel vs undocumented ImageCases. $\endgroup$ – user5601 Feb 27 '19 at 13:46
  • 2
    $\begingroup$ ImageCases will be upgraded in v12 apparently. I asked in one of the livestreams, and apparently it uses some variant of YOLO under the hood, so I wouldn't necessarily expect major gains if it doesn't currently work for you. $\endgroup$ – Carl Lange Feb 27 '19 at 14:59
2
$\begingroup$

In version 12, functions such as ImageCases and ImageBoundingBoxes exist that will do this job better than FindFaces. I'm pretty certain it is using YOLO under the hood - I'm not sure any other network makes sense.

Anyway, here is how you use it:

i = Import["http://img.burrard-lucas.com/nepal/full/bhaktapur_street_scene.jpg"];
ImageBoundingBoxes[i, Entity["Concept", "Person::93r37"], AcceptanceThreshold -> .2] // HighlightImage[i, #]&

enter image description here

This will at least do a better job than FindFaces would (not least because not everybody is facing the camera at all times ;) )

In pre-12 versions, you can create this functionality yourself using the YOLO v2 network directly.

|improve this answer
$\endgroup$

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.