search

chungyiweng / humannerf

HumanNeRF turns a monocular video of moving people into a 360 free-viewpoint video.
MIT License
786 stars 86 forks source link

Getting weird results when training with custom data #50

Open qzane opened 1 year ago

qzane commented 1 year ago

prog_025000 I got this weird result, is it normal or did I do any wrong? I validate my camera parameters using cv2.projectPoints and they seems to be correct. Also, I have another dataset in which I got the body pose using EasyMocap, can I just use the 'pose' and 'Rh' data from them (I also checked the code in prepare_dataset.py and find that you actually don't need 'Th', right?

Dipankar1997161 commented 1 year ago

@qzane,

Even I am getting weird results for my dataset when training. Did you generate your smpl values from ROMP/VIBE?

Maybe its the camera values or something wrong. If you solved it, it would be great if you can provide some information on this.

I will be waiting for your response

prog_005000

leviome commented 1 year ago

LOL, it seems I was not the only one getting this: prog_040000

StephenYangjz commented 1 year ago

I am experiencing something similar with the results form 4D Humans. Has anyone solved this issue?

Dipankar1997161 commented 1 year ago

I am experiencing something similar with the results form 4D Humans. Has anyone solved this issue?

That's because the projection of sampling rays isn't 100% accurate. So make sure you have the correct camera values.

StephenYangjz commented 1 year ago

I am experiencing something similar with the results form 4D Humans. Has anyone solved this issue?

That's because the projection of sampling rays isn't 100% accurate. So make sure you have the correct camera values.

I ran 4d humans on mocap just to verify, and I even tried to substitute the cam in/extrinsics by the mocap dataset gt. It still gave me similar results...

Dipankar1997161 commented 1 year ago

I am experiencing something similar with the results form 4D Humans. Has anyone solved this issue?

That's because the projection of sampling rays isn't 100% accurate. So make sure you have the correct camera values.

I ran 4d humans on mocap just to verify, and I even tried to substitute the cam in/extrinsics by the mocap dataset gt. It still gave me similar results...

What do you mean by running mocap on it? Did you mean Easymocap?