Testing PyTorch implementation on my own images

[kvnsng]

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:

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]

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]

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

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

[xingyizhou]

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]

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]

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

[Ariel-JUAN]

@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?