search

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
30.86k stars 7.83k forks source link

Inference Time difference for multi-person vs single-person #867

Closed vikasgupta-github closed 5 years ago

vikasgupta-github commented 5 years ago

In the readme, it is written that for Body Pose estimation, the inference time is independent of the number of persons.

When I run the code for 1 or 2 persons, the difference is negligible. But when I compare with 1 person vs 9 person, then the inference time becomes 1.5 times that of single person.

Given below are the images I have used. NOTE : The images are of same size.

000000000785

000000044279

000000016309

gineshidalgo99 commented 5 years ago

I can run 20 people at about 22FPS in a Titan Xp while 1 person is about 23-24 FPS, so the difference is less than 10%. Something else is going on in your particular case. If you want me to give you a possible higher insight to your particular case, please post the hardward config as our GitHub issue template indicates. Thanks.