CMU-Perceptual-Computing-Lab / openpose

OpenPose: Real-time multi-person keypoint detection library for body, face, hands, and foot estimation
https://cmu-perceptual-computing-lab.github.io/openpose
Other
31.41k stars 7.88k forks source link

Solution for Foot-only Keypoint Detection #1203

Closed ChenFengYe closed 5 years ago

ChenFengYe commented 5 years ago

Hi, thank for your wonderful work. Foot estimation is really helpful.

However, I find the foot part will lose, even if the foot is obvious in the center of the image(not a complete person). I guess Oposepose focuses on the person first, so is there any solution for foot-only keypoint detection?

gineshidalgo99 commented 5 years ago

Updated: The solution to detect standalone feet is to use --maximize_positives, read the doc/demo_overview.md for more details.

Explanation The reason is that it automatically skip the following 3 lines of code in https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/src/openpose/net/bodyPartConnectorBase.cpp#L641

                    // No consider foot keypoints for that
                    for (auto i = 19 ; i < 25 ; i++)
                        personCounter -= (peopleVector[index].first.at(i) > 0);

That way, keypoints from foot-only images should be printed as well. Let me know if it works for your case. It increases the number of false positives, that is why we do this, but if you want foot-only pics, this is the way to make it work.

ChenFengYe commented 4 years ago

@gineshidalgo99 Thanks for your reply. Nice work!

emirhanylmzz commented 3 years ago

Updated: The solution to detect standalone feet is to use --maximize_positives, read the doc/demo_overview.md for more details.

Explanation The reason is that it automatically skip the following 3 lines of code in https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/src/openpose/net/bodyPartConnectorBase.cpp#L641

                    // No consider foot keypoints for that
                    for (auto i = 19 ; i < 25 ; i++)
                        personCounter -= (peopleVector[index].first.at(i) > 0);

That way, keypoints from foot-only images should be printed as well. Let me know if it works for your case. It increases the number of false positives, that is why we do this, but if you want foot-only pics, this is the way to make it work.

I understand you and tried your solution but the result accuracy is too bad. Is there any way to train only foot database and detect only foot keypoints on foot-only pics.

PeerW commented 3 years ago

@emirhanylmzz Did you find a solution? I'm experiencing the same low accuracy and would also be interested in a way to detect only foot key points and train on only foot images.

emirhanylmzz commented 3 years ago

@emirhanylmzz Did you find a solution? I'm experiencing the same low accuracy and would also be interested in a way to detect only foot key points and train on only foot images.

No, sorry :(

devendraswamy commented 2 years ago

Guys have you trained the foot keypoint detection , if you done then kindly help me. if possible please share the your work or open source links here , I am very thankful to you in advance.