search

xingyizhou / pytorch-pose-hg-3d

PyTorch implementation for 3D human pose estimation
GNU General Public License v3.0
613 stars 141 forks source link

Testing PyTorch implementation on my own images #63

Open kvnsng opened 5 years ago

kvnsng commented 5 years ago

Hi, thanks for open sourcing this great work! I'm looking forward to trying and using it further in the future.

I've tried running the demo.py on just a random CG human image I found on google, but I get very bad results:

image

Do you have any idea why it might be performing badly on this image, and is there a way to make it work?

Thanks!

xingyizhou commented 5 years ago

Hi kevin, Thanks for your interests in our work. This is expected. Our method does not work on such images. The major reason is that our method requires complete human as input, not cropped half body, since the model is trained only on compele human images. You will need to trained the model and add half body crop data augmentation to fix this.

kvnsng commented 5 years ago

oh, I'm sorry. The image was just a screenshot, and I ended up cropping out the bottom half. here's the full image.

image

I ran the model on this image, and it gave me the 3d results shown above... Any idea what could be happening?

xingyizhou commented 5 years ago

Hi Keven, Thanks for the report! This is a typo of my demo script that causes the input being not resized before feeding to the network. I have fixed it. The current version should work fine.

zhengyuezhi commented 5 years ago

Hi kevin, Thanks for your interests in our work. This is expected. Our method does not work on such images. The major reason is that our method requires complete human as input, not cropped half body, since the model is trained only on compele human images. You will need to trained the model and add half body crop data augmentation to fix this.

Hello, as you mentioned, when you add a half-cutting data enhancement to the training model to solve the problem, then this half-cut data is made into MPII format or human3.6 format data set. If it is human3.6, how should it be marked?

kvnsng commented 5 years ago

Okay, great! I'll give it a shot today and let you know how it goes. Thanks so much for your help!

Ariel-JUAN commented 5 years ago

@xingyizhou Hi, I found that when people in the middle of image, the pose estimation works fine. But if people not in the middle of image, the result is terrible. Do you notice that? How should I work on this problem in order to works all situations if people not in the middle of image?