Openpose: Demo very slow

Created on 16 Aug 2018  路  12Comments  路  Source: CMU-Perceptual-Computing-Lab/openpose

Issue Summary

Demo very slow after following the installing tutorial

Executed Command (if any)

Note: add --logging_level 0 --disable_multi_thread to get higher debug information.

./build/examples/openpose/openpose.bin --video examples/media/video.avi

OpenPose Output (if any)

a window with images but at really slow rate, it's never been greater than 0 fps...

Type of Issue

You might select multiple topics, delete the rest:

  • Compilation/installation error
    As it was noticed, I unchecked the box to avoid installing Caffe in cmake-gui. Before I used brew to install it, then this step passed!
  • Execution error
  • Help wanted
    I saw that a problem could be resolved with #423 by cloning caffe manually to pass the build stage.
    First, by cloning openpose the directory caffe exists... and in fact, the problem at build stays.
  • Question
    I suppose the problem comes from caffe and its installing using brew (https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/doc/installation.md#mac-osx-version) as openpose doesn't work properly.
    So can you help me to build Caffe within cmake-gui?
  • Enhancement / offering possible extensions / pull request / etc
    Harmonization between #423 and (https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/doc/installation.md#mac-osx-version)
  • Other (type your own type)

Your System Configuration

  1. OpenPose version: Latest GitHub code? Or specific commit (e.g., d52878f)? Or specific version from Release section (e.g., 1.2.0)?
    1.3.0

  2. General configuration:

    • Installation mode: CMake, sh script, manual Makefile installation, ... (Ubuntu); CMake, ... (Windows); ...?
      cmake-gui
    • Operating system (lsb_release -a in Ubuntu): macOS High Sierra 10.13.6
    • Release or Debug mode? (by default: release):
    • Compiler (gcc --version in Ubuntu or VS version in Windows): 5.4.0, ... (Ubuntu); VS2015
      Enterprise Update 3, VS2017 community, ... (Windows); ...?
      I launched make -j'nproc'
  3. Non-default settings:

    • 3-D Reconstruction module added? (by default: no):
    • Any other custom CMake configuration with respect to the default version? (by default: no):
  4. 3rd-party software:

    • Caffe version: Default from OpenPose, custom version, ...?
      Default from brew
    • CMake version (cmake --version in Ubuntu):
    • OpenCV version: pre-compiled apt-get install libopencv-dev (only Ubuntu); OpenPose default (only Windows); compiled from source? If so, 2.4.9, 2.4.12, 3.1, 3.2?; ...?

3.4.2

  1. If GPU mode issue:

    • CUDA version (cat /usr/local/cuda/version.txt in most cases):
    • cuDNN version:
    • GPU model (nvidia-smi in Ubuntu):
  2. If CPU-only mode issue: none

  3. If Python API:

    • Python version: 2.7, 3.7, ...? 3
    • Numpy version (python -c "import numpy; print numpy.version.version" in Ubuntu):
  4. If Windows system:

    • Portable demo or compiled library?
  5. If speed performance issue:

    • Report OpenPose timing speed based on this link.

0 fps

help wantequestion

Most helpful comment

Guys I'm using a GTX 1080 and manage to get 45/50fps on other pose estimation systems based on Cuda but only up to 12fps on OpenPose. Is this expected or am I doing something wrong?

Please can you tell me which one did you use i am faceing the same problem with openPose very slow and i have no idea how to make it faster or find other library

All 12 comments

Could you try COCO instead of BODY_25? BODY_25 seems to be super slow in CPU version.

Hello,

I generate the build once again with the download option COCO set, then compile.
To try it I launched ./build/examples/openpose/openpose.bin -model_pose='COCO' --video examples/media/video.avi
But there were no better results (very slow).

The expected speed (shown in the OpenPose doc) of that one is about 0.3 FPS (COCO), rather than 0.1 FPS (BODY_25). Which speed are you getting?

Guys I'm using a GTX 1080 and manage to get 45/50fps on other pose estimation systems based on Cuda but only up to 12fps on OpenPose. Is this expected or am I doing something wrong?

That's expected. If you desire faster speed, reduce the net_resolution (check the doc for details)

Hi @gineshidalgo99 ,

I made a launch of the video demo with profiling speed flag, in order to have more accurate measures:
macbook-pro-de-sebastien-2:openpose sebast$ ./build/examples/openpose/openpose.bin --profile_speed=12 --video examples/media/video.avi
Starting OpenPose demo...
Starting thread(s)...
31.062490 msec at in /Users/sebastien 1/development/openpose/include/openpose/producer/wDatumProducer.hpp:workProducer():70
0.001030 msec at in /Users/sebastien 1/development/openpose/include/openpose/thread/wIdGenerator.hpp:work():69
0.004955 msec at in /Users/sebastien 1/development/openpose/include/openpose/core/wScaleAndSizeExtractor.hpp:work():67
6.691391 msec at in /Users/sebastien 1/development/openpose/include/openpose/core/wCvMatToOpInput.hpp:work():64
24.076623 msec at in /Users/sebastien 1/development/openpose/include/openpose/core/wCvMatToOpOutput.hpp:work():66
36332.615372 msec at in /Users/sebastien 1/development/openpose/include/openpose/pose/wPoseExtractor.hpp:work():91
20.803803 msec at in /Users/sebastien 1/development/openpose/include/openpose/pose/wPoseRenderer.hpp:work():72
23.449569 msec at in /Users/sebastien 1/development/openpose/include/openpose/core/wOpOutputToCvMat.hpp:work():62
0.899657 msec at in /Users/sebastien 1/development/openpose/include/openpose/gui/wGuiInfoAdder.hpp:work():65
79.902224 msec at in /Users/sebastien 1/development/openpose/include/openpose/gui/wGui.hpp:workConsumer():80

Please notice the duration of wPoseExtractor.hpp:work() thread of 36 seconds...
Thx in advance

That's normal for BODY_25, but the COCO model should be less than 1 second. Could you confirm this? Thanks. (BODY_25 is faster in GPU than COCO, but COCO is much much faster than BODY_25 in CPU version)

With COCO not faster than a frame every 20 seconds.
./build/examples/openpose/openpose.bin --profile_speed 10 --model_pose COCO --video examples/media/video.avi
Starting OpenPose demo...
Starting thread(s)...
^C

bye

Sorry, I did not realize you are in Mac.

MKL-Caffe (CPU version) provides about 0.3 FPS in Ubuntu, but non-MKL Caffe (in both CPU Ubuntu and Mac) are as slow as you are pointing out. So it's the normal speed.

I think MKL Caffe is not supported in Mac. If you were able to install MKL-Caffe, and link OpenPose against it, then you should solve this speed issue and go up to about 0.1-0.3 FPS.

With COCO not faster than a frame every 20 seconds.
./build/examples/openpose/openpose.bin --profile_speed 10 --model_pose COCO --video examples/media/video.avi
Starting OpenPose demo...
Starting thread(s)...
^C

bye

A picture takes 50 seconds on my computer.

@yavana If you look for help, please include your hw configuration as the GitHub issue template indicates.

Guys I'm using a GTX 1080 and manage to get 45/50fps on other pose estimation systems based on Cuda but only up to 12fps on OpenPose. Is this expected or am I doing something wrong?

Please can you tell me which one did you use i am faceing the same problem with openPose very slow and i have no idea how to make it faster or find other library

Was this page helpful?
0 / 5 - 0 ratings