demo.py cannot show result

[ujsyehao]

when I run python examples/demo.py, the program can work normally: but cannot show the result, Figure 1 is empty:

I have some extra information to provide:

• run "make runtest -j8", it has an error PC: @ 0x7f0d5cb2157f caffe::SpatialTransformerLayer<>::Backward_gpu(), I investigate the error, the implementation of SpatialTransformer layer cause the problem.

• run "jupyter notebook examples/rmpe/Regional\ Multi-person\ Pose\ Estimation.ipynb", it has error the server cannot work, the problem is my system problem, I cannot use jupyter.

Thank you in advance @Fang-Haoshu

[Fang-Haoshu]

It's weird...Can you try to store the image instead of drawing it? Like using plt.savefig()

[ujsyehao]

Hi, when I add plt.savefig(path) after plt.draw(), it can show the correct result, but I do not know the reason. Can you tell me the reason? Thank you in advance.

I have another confusion, I test the speed use sample images(demo.py provide): average speed is 0.74 second but in Read.md

1.5s/frame

So I am confused it.

Is there any problem with my test method? Below is my test method code: start = time.time() image = caffe.io.load_image(filename) ... for i in xrange(len(preds)): __... end = time.time() print end - start plt.draw() plt.savefig('mypath') time.sleep(1)

[ujsyehao]

Hi, @Fang-Haoshu Do you have time to solve the problem? Thank you in advance!

[Fang-Haoshu]

Hi, this is not problem. The time varies according to the number of people per image.

[ujsyehao]

I am sorry to bother you again, I am a newbie to pose estimation field and cannot figure out the time cost in RMPE/AlphaPose program.

As far as I know, use one image to test speed(only discuss batchsie=1 situation), time cost = detection time + pose estimation time No matter how many people in a picture, it brings slight time cost difference in object detection(faster RCNN RPN network always output 300 proposals, then classify network operates on these 300 proposals), However, you say

The inference time depends on the number of persons in the image. On a Titan X GPU card, the inference time is about 0.2s per person.

So 0.2s = per person pose estimation time without people detection or 0.2s = per person pose estimation time plus people detection time?

[Fang-Haoshu]

human detection is fast and thus the average time for each person can be ignored.

[ujsyehao]

@Fang-Haoshu So 0.2s = per person pose estimation time without people detection?