@@ -78,4 +78,4 @@ You might select multiple topics, delete the rest:
...
@@ -78,4 +78,4 @@ You might select multiple topics, delete the rest:
- Portable demo or compiled library?
- Portable demo or compiled library?
10. If **speed performance** issue:
10. If **speed performance** issue:
- Report OpenPose timing speed based on [this link](https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/doc/installation.md#profiling-speed).
- Report OpenPose timing speed based on [this link](https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/doc/speed_up_openpose.md#profiling-speed).
@@ -158,7 +158,7 @@ Output (format, keypoint index ordering, etc.) in [doc/output.md](doc/output.md)
...
@@ -158,7 +158,7 @@ Output (format, keypoint index ordering, etc.) in [doc/output.md](doc/output.md)
## Speeding Up OpenPose and Benchmark
## Speeding Up OpenPose and Benchmark
Check the OpenPose Benchmark as well as some hints to speed up and/or reduce the memory requirements for OpenPose on [doc/speed_up_preserving_accuracy.md](doc/speed_up_preserving_accuracy.md).
Check the OpenPose Benchmark as well as some hints to speed up and/or reduce the memory requirements for OpenPose on [doc/speed_up_openpose.md](doc/speed_up_openpose.md).
**Q: Low speed** - OpenPose is quite slow, is it normal? How can I speed it up?
**Q: Low speed** - OpenPose is quite slow, is it normal? How can I speed it up?
**A**: Check [doc/speed_up_preserving_accuracy.md](./speed_up_preserving_accuracy.md) to discover the approximate speed of your graphics card and some speed tips.
**A**: Check [doc/speed_up_openpose.md](./speed_up_openpose.md) to discover the approximate speed of your graphics card and some speed tips.
### CPU Version Too Slow
### CPU Version Too Slow
**Q: The CPU version is insanely slow compared to the GPU version.**
**Q: The CPU version is insanely slow compared to the GPU version.**
**A**: Check [doc/speed_up_preserving_accuracy.md#cpu-version](./speed_up_preserving_accuracy.md#cpu-version) to discover the approximate speed and some speed tips.
**A**: Check [doc/speed_up_openpose.md#cpu-version](./speed_up_openpose.md#cpu-version) to discover the approximate speed and some speed tips.
### Profiling Speed and Estimating FPS without Display
### Profiling Speed and Estimating FPS without Display
Check the [doc/installation.md#profiling-speed](./installation.md#profiling-speed) section.
Check the [doc/speed_up_openpose.md#profiling-speed](./speed_up_openpose.md#profiling-speed) section.
...
@@ -109,7 +109,7 @@ COCO model will eventually be removed. BODY_25 model is faster, more accurate, a
...
@@ -109,7 +109,7 @@ COCO model will eventually be removed. BODY_25 model is faster, more accurate, a
### How to Measure the Latency Time?
### How to Measure the Latency Time?
**Q: How to measure/calculate/estimate the latency/lag time?**
**Q: How to measure/calculate/estimate the latency/lag time?**
**A**: [Profile](https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/doc/installation.md#profiling-speed) the OpenPose speed. For 1-GPU or CPU-only systems (use `--disable_multi_thread` for simplicity in multi-GPU systems for latency measurement), the latency will be roughly the sum of all the reported measurements.
**A**: [Profile](https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/doc/speed_up_openpose.md#profiling-speed) the OpenPose speed. For 1-GPU or CPU-only systems (use `--disable_multi_thread` for simplicity in multi-GPU systems for latency measurement), the latency will be roughly the sum of all the reported measurements.
@@ -4,9 +4,9 @@ OpenPose - Maximizing the OpenPose Speed
...
@@ -4,9 +4,9 @@ OpenPose - Maximizing the OpenPose Speed
## Contents
## Contents
1.[OpenPose Benchmark](#openpose-benchmark)
1.[OpenPose Benchmark](#openpose-benchmark)
2.[Profiling Speed](#profiling-speed)
2.[Profiling Speed](#profiling-speed)
3.[Speed Up Preserving Accuracy](#speed-up-preserving-accuracy)
3.[CPU Version](#cpu-version)
4.[Speed Up and Memory Reduction](#speed-up-and-memory-reduction)
4.[Speed Up Preserving Accuracy](#speed-up-preserving-accuracy)
5.[CPU Version](#cpu-version)
5.[Speed Up and Memory Reduction](#speed-up-and-memory-reduction)
...
@@ -17,18 +17,28 @@ Check the [OpenPose Benchmark](https://docs.google.com/spreadsheets/d/1-DynFGvoS
...
@@ -17,18 +17,28 @@ Check the [OpenPose Benchmark](https://docs.google.com/spreadsheets/d/1-DynFGvoS
### CPU Version
The CPU version runs at about 0.3 FPS on the COCO model, and at about 0.1 FPS (i.e., about 15 sec / frame) on the default BODY_25 model. Switch to COCO model and/or reduce the `net_resolution` as indicated above. Contradictory fact: BODY_25 model is about 5x slower than COCO on CPU-only version, but it is about 40% faster on GPU version.
### Profiling Speed
### Profiling Speed
Check the [doc/installation.md#profiling-speed](./installation.md#profiling-speed) section to measure the bottlenecks in your OpenPose distribution and make sure everything is working as expected.
OpenPose displays the FPS in the basic GUI. However, more complex speed metrics can be obtained from the command line while running OpenPose. In order to obtain those, compile OpenPose with the `PROFILER_ENABLED` flag on CMake-gui. OpenPose will automatically display time measurements for each subthread after processing `F` frames (by default `F = 1000`, but it can be modified with the `--profile_speed` flag, e.g. `--profile_speed 100`).
- Time measurement for 1 graphic card: The FPS will be the slowest time displayed in your terminal command line (as OpenPose is multi-threaded). Times are in milliseconds, so `FPS = 1000/millisecond_measurement`.
- Time measurement for >1 graphic cards: Assuming `n` graphic cards, you will have to wait up to `n` x `F` frames to visualize each graphic card speed (as the frames are splitted among them). In addition, the FPS would be: `FPS = minFPS(speed_per_GPU/n, worst_time_measurement_other_than_GPUs)`. For < 4 GPUs, this is usually `FPS = speed_per_GPU/n`.
Make sure that `wPoseExtractor` time is the slowest timing. Otherwise the input producer (video/webcam codecs issues with OpenCV, images too big, etc.) or the GUI display (use OpenGL support as detailed in [doc/speed_up_openpose.md](./speed_up_openpose.md)) might not be optimized.
## Speed Up Preserving Accuracy
## Speed Up Preserving Accuracy
Some speed tips to maximize the OpenPose runtime speed while preserving the accuracy (do not expect miracles, but it might help a bit boosting the framerate):
Some speed tips to maximize the OpenPose runtime speed while preserving the accuracy (do not expect miracles, but it might help a bit boosting the framerate):
1. Enable the `WITH_OPENCV_WITH_OPENGL` flag in CMake to have a much faster GUI display (but you must also compile OpenCV with OpenGL support). Note: Default OpenCV in Ubuntu (from apt-get install) does have OpenGL support included. Nevertheless, default Windows portable binaries do not.
1. Enable the `WITH_OPENCV_WITH_OPENGL` flag in CMake to have a much faster GUI display. It reduces the lag and increase the speed of displaying images by telling OpenCV to render the images using OpenGL support. This speeds up display rendering about 3x. E.g., it reduces from about 30 msec to about 3-10 msec the display time for HD resolution images. It requires OpenCV to be compiled with OpenGL support and it provokes a visual aspect-ratio artifact when rendering a folder with images of different resolutions. Note: Default OpenCV in Ubuntu 16 (from apt-get install) does have OpenGL support included. Nevertheless, default one from Ubuntu 18 and the Windows portable binaries do not.
2. Change GPU rendering by CPU rendering to get approximately +0.5 FPS (`--render_pose 1`).
2. Change GPU rendering by CPU rendering to get approximately +0.5 FPS (`--render_pose 1`).
3. Use cuDNN 5.1 (cuDNN 6 is ~10% slower).
3. Use cuDNN 5.1 or 7.2 (cuDNN 6 is ~10% slower).
4. Use the `BODY_25` model for simultaneously maximum speed and accuracy (both COCO and MPII models are slower and less accurate).
4. Use the `BODY_25` model for simultaneously maximum speed and accuracy (both COCO and MPII models are slower and less accurate). But it does increase the GPU memory, so it might go out of memory more easily in low-memory GPUs.
...
@@ -39,8 +49,3 @@ Some speed tips to highly maximize the OpenPose speed, but keep in mind the accu
...
@@ -39,8 +49,3 @@ Some speed tips to highly maximize the OpenPose speed, but keep in mind the accu
2. For face, reduce the `--face_net_resolution`. The resolution 320x320 usually works pretty decently.
2. For face, reduce the `--face_net_resolution`. The resolution 320x320 usually works pretty decently.
3. Points 1-2 will also reduce the GPU memory usage (or RAM memory for CPU version).
3. Points 1-2 will also reduce the GPU memory usage (or RAM memory for CPU version).
4. Use the `BODY_25` model for maximum speed. Use `MPI_4_layers` model for minimum GPU memory usage (but lower accuracy, speed, and number of parts).
4. Use the `BODY_25` model for maximum speed. Use `MPI_4_layers` model for minimum GPU memory usage (but lower accuracy, speed, and number of parts).
### CPU Version
The CPU version runs at about 0.3 FPS on the COCO model, and at about 0.1 FPS (i.e., about 15 sec / frame) on the default BODY_25 model. Switch to COCO model and/or reduce the `net_resolution` as indicated above. Contradictory fact: BODY_25 model is about 5x slower than COCO on CPU-only version, but it is about 40% faster on GPU version.
