Bent knees of reconstructed model!!!

[vchoutas]

Bent knees of reconstructed model!!!

[Anirudh257]

@Tanupvats Can you post your input image? Is it having occlusion?

[Tanupvats]

sorry for replying late @Anirudh257 and thanks for asking.

[Anirudh257]

@Tanupvats No problem. Can you also post the RGB image on which you have run SMPL-X on?

[Tanupvats]

@Anirudh257 ya sure! here is the image

[Anirudh257]

I am able to get a proper image with the knees straight.

What were your hyperparameters? These are the parameters that I used:

python3 smplifyx/main.py --config cfg_files/fit_smplx.yaml --data_folder data_folder_check/ --output_folder output_folder_check/ --visualize=False --model_folder model_folder/models/ --vposer_ckpt vposer_folder/vposer_v1_0/ --interpenetration False --focal_length 1500

[Tanupvats]

@Anirudh257 actually I tried to run the image on the Inference.py file they have provided and used the same parameters as they did. I am a new bee in this field and I don't
have much knowledge about this. so could you please give me instruction or some sort of follow up path to achieve this result?.
And Thank you so much for replying.

[Anirudh257]

@Tanupvats My bad!! I used https://github.com/vchoutas/smplify-x/ to run the code. You can check it out once. Follow the instructions as given by the authors on their GitHub repository and you will get the result. Do extract the openpose keypoints as well.

[Tanupvats]

hi, @Anirudh257 I managed to run get the results using the same parameters as you did but I got this.

could you tell me why this is happening?

[Anirudh257]

It doesn't have bent knees. It seems to be correct. What is the exact issue?

[Tanupvats]

@Anirudh257 I mean it's tilted towards its front while the person is standing straight.

[Anirudh257]

Okay, I am not so sure, but you can try changing the focal length parameters and other parameters. If you feel that the results don't improve, try using https://github.com/ahmedosman/STAR, https://github.com/vchoutas/smplify-x or other SMPL-X variants.