saurabhsharma1993 / generative_pose Goto Github PK

hi,
thanks for sharing the implementation.
I have a question about format of data. As I understood, train_2d_ft_norm is a dictionary that its keys are about subjects, act, and videos of h36m dataset and the values have a format of number of images * 32. Can you tell me that where does "32" come from? I think that it is related to joints but we have 17 joints in h36m. If i am right, can you tell me that which 16 joints does it referring to?

hi:
When I finished training the network and tested it on the test data set, I found that the result was very different from that in the paper, and I did not change any parameters of the source code，So what's the reason? @ssfootball04

May I ask a silly quetsion... What's the meaning of oracle?

My understanding of oracle-based pose sample is that, for the output of CVAE, you will select the one which has the nearest distance to the GT data.

But why do you call this using an Oracle that has access to P_{3D}?
Can we do this during test time (I mean a real-world scene, no 3D gt data)?

Thanks for great work
I run the following line:
python main.py --exp example--test --numSamples 10 --load 'best_model.pth.pt'

the results are all numbers. and results/test folder was created but empty, is there any visualization such as images with 3d pose estimation

Hi! Thanks for the nice work! If I understand it correctly, the released code only contains the CVAE part of the ICCV paper? Could I find the Ordinal Net somewhere?

Thanks for sharing your work!

I have a question about your training details. Do you use CMU Mocap dataset to trian your model when using 2D-3D paired data for evaluation?

Thanks for your fantastic work!
I have a question about cvae training. I find KL divergence loss decreased to a very small value(about 1e-3, some times it decreased to 0) In VAE, network may suffer from KL-vanishing. Will CVAE suffer from this same problem？What is the common KL loss value in network training？
Looking for your reply, thanks!

How to visualize the images using viz.py file?

Hello! I wonder whether you have ever tried generating the best 3D pose according to the ordinal score map directly? And Why not score the 3D pose according to the distance between prediction and GT with generated multi-pose directly?