CPU mode gives wrong 3D pose results

Question

CPU mode gives wrong 3D pose results

handsomeyang opened this issue 7 years ago · 18 comments

python demo_tf.py --device=cpu \
		   --demo_type=image \
		   --test_file=test_imgs/yuniko.jpg \
		   --model_file=models/weights/vnect_tf \
		   --plot_2d=True \
		   --plot_3d=True

Answer 1 · 2017-11-30T15:54:07.000Z

I think it's not related to cpu or gpu, but there's some problem in 3d pose now.

Answer 2 · 2017-12-04T13:22:12.000Z

Has anyone traced the cause of the error?

Answer 3 · 2017-12-04T13:41:03.000Z

I am taking same results. Has anyone solved this problem ?

Answer 4 · 2017-12-04T14:14:26.000Z

@timctho I think there is some problem in 2d pose results. Joints are not in true localizations.

Answer 5 · 2017-12-26T09:53:23.000Z

can any one send me the VNect model file? Thanks!

Answer 6 · 2018-01-20T10:31:48.000Z

@timctho How can I get the models/weights/vnect_tf file?

Answer 7 · 2018-01-28T20:16:24.000Z

Mabe, it just missed the step of "Kinematic Skeleton Fiting”

Answer 8 · 2018-01-29T02:31:14.000Z

@Superlee506 can you tell me more about "Kinematic Skeleton Fiting”? Is it wrote in the paper?

Answer 9 · 2018-05-07T10:40:17.000Z

3d pose is also a problem, i test some but cant get a good result, anybody has good idea?

Answer 10 · 2018-05-15T16:38:47.000Z

there is a mistake in utils/utils.py

in def extract_3d_joints_from_heatmap x and y is mixed up (line 80 and 81). it should be

coord_2d_x = joints_2d[joint_num][0] 
coord_2d_y = joints_2d[joint_num][1]

After fixing that the 3D plot looks fine (although the axes are messed up)

Answer 11 · 2018-06-14T07:02:27.000Z

This is my output from above method.
It is quite different from above result.
In my case, 3D skeleton is also wrongly estimated.
How can I get the good estimation?

Answer 12 · 2018-06-14T09:10:14.000Z

Well not even your 2D pose is correct and it should be working if you checkout the current repository.
Are sure you changed the correct lines of code?

Answer 13 · 2018-06-14T10:42:44.000Z

@cDenius
I had tried this code with these follows.

Download weights from the author of paper
change weights with using './caffe_weight_to_pickle.py'
run '/models/vnect_model.py' to get

   vnect_tf.data-00000-of-00001
   vnect_tf.index
   vnect_tf.meta

change the code

   coord_2d_x = joints_2d[joint_num][0] 
   coord_2d_y = joints_2d[joint_num][1]

run '/demo_tf.py'

If you can share your parameters, could you send it to my e-mail?

powermew@hanyang.ac.kr

Answer 14 · 2018-06-14T14:45:48.000Z

That is weird. If you don't change the two lines of code do you get the same results as the OP?
I get the desired output without doing or changing anything else besides the 5 steps you have described.

Answer 15 · 2018-06-14T15:06:24.000Z

@cDenius
So do you have tensorflow weights without conversion??
Where did you get that weights?
i mean vnect_tf~ files

Answer 16 · 2018-07-02T11:09:22.000Z

The messed up axes can be fixed by switching y_pair and x_pair in draw_limbs_3d in utils.py (lines 112+113):

Although there seem to be some scaling issues now haha

Answer 17 · 2018-10-09T08:56:01.000Z

@lavage
I can reproduce your results by following your procedures, but it seems like the z coordinates of 3d joint points are incorrect.
Have you observed any similar situations?
Thanks.

Answer 18 · 2018-10-09T09:55:10.000Z

@zeroHuang0516
That is what I meant with the scaling issue, you just defined it a bit better ;). Unless you are having different z coordinate issues than shown in my picture above?

As the overall body composition looks ok, it wouldn't be too difficult to normalize the z coordinates however.