About the running time

[lianuo]

@TreB1eN Thank you for your works, it is easy to test the openpose method in pytorch.

I have a question about the running speed.

I use this model to test a video with 10 person in each frame.it seems too slow (7 second per frame) compare with the paper they reported in the fig 12 as follow:

could you give me some information I could have faster speed for inference.

I am using the precise mode.

Thank you~

[TreB1eN]

firstly, could you try not to use the precise mode? Because in precise mode the inference will be running multiple times, each time on a different input scale. Also, the inference time are also dependent on your GPU efficiency. On my 1080Ti one inference time is less than 0.5 seconds, as I remembered. Right now my resource are computing on other tasks, If I got time, I may give you a precise figure

[lianuo]

@TreB1eN Thank you for reply, I test the time spend ,and found most of time spend on 1 compute_peaks_from_heatmaps 2 compute_connections

I fix the scale to 2 ,remove 0.5 1 1.5

when i resize my image resolution from 1920/1080->576/324 the time of 1 is down 1.8598 -> 0.1488 s and 2 down to 3.3724->1.7937 s

do you think we can improve the code to be faster?

[TreB1eN]

really ? What I recall the step2 is pretty fast, I may need a while for me to test it again in my computer

[lianuo]

@TreB1eN I think it is sensitive to the person or key points number in the frames.in my image there are 10 or more persons.so it is very slow.

[lianuo]

看了下代码，heatmap和PAF之后的处理这一部分是在CPU里面实现的，而openpose源码是在cuda下面跑的，大概跟这个有些关系吧。

[TreB1eN]

模型加速是真的很难，不过有个lightweight openpose

[lianuo]

@TreB1eN 谢谢你提供的这个，我看他也开源了，测了一下，发现有相同的问题，就是计算极大值和关键点连接是最耗时的部分。计算方法和你这个方法也很接近。

[TreB1eN]

@lianuo 如果想加速，还是要用c实现，或者Gpu

[yy2yy]

谢谢你提供的这个，请问你的handnet模型是在哪里找的？是自己训练的吗？用的什么数据？