dataset about People_SnapShot

[JiatengLiu]

Hello!! Have you test your framework on dataset People_SnapShot?

[skhu101]

No, we did not test it on People_SnapShot data set.

[JiatengLiu]

Thanks for your reply, and I want to test it on people_snapshot, do you know how to process data sets to fit GauHuman?

[skhu101]

Hi, maybe one quick way is to follow the data loader used in GauHuman.

[JiatengLiu]

Thanks for your reply, I have tried like this but some parameters were wrong. such as theta in smpl params is not zero in people_snapshot, which is zero in zju_mocap. I tried to assign them directly to zero and got what plausible result like below. But you know, deleting data from a data set is not an appropriate thing to do.

[skhu101]

Hi, I would suggest you check by projecting SMPL vertices to the image space. Try to make sure that SMPL parameters are correctly loaded.

[JiatengLiu]

Thanks for your suggest! I have fix it:>     ------------------ Original ------------------ From: @.>; Date:  Thu, May 16, 2024 09:34 AM To: @.>; Cc: @.>; @.>; Subject:  Re: [skhu101/GauHuman] dataset about People_SnapShot (Issue #32)

 

Hi, I would suggest you check by projecting SMPL vertices to the image space. Try to make sure that SMPL parameters are correctly loaded.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

[yashgarg98]

Hi @skhu101 and @JiatengLiu, can you let me know how you created/preprocessed the files to create the dataloader for the people_snapshot dataset? I have followed the https://github.com/zju3dv/instant-nvr/blob/master/docs/preprocess.md to create the files but I'm assuming some files are still missing.

Thanks

[JiatengLiu]

Hi @yashgarg98, Just follow this repo, processing datasets in this repo is suitable for GauHuman's needs. https://github.com/wenbin-lin/RelightableAvatar

[Yi-Cheng0101]

Hi @JiatengLiu, I follow the repo you provided. But, they still missed some files, such as annots.npy and smpl_params. Could you share how you get them through the repo you shared. Thanks!

[JiatengLiu]

My directory looks like this, using male-3-casual as an example.

├── camera.pkl
├── consensus.obj
├── consensus.pkl
├── image
├── keypoints.hdf5
├── lbs
├── male-3-casual.mp4
├── mask
├── masks.hdf5
├── params.npy
├── reconstructed_poses.hdf5
├── tex-male-3-casual.jpg
└── vertices

If you need, I can share my dataset with you:> @Yi-Cheng0101

[Yi-Cheng0101]

Hi @JiatengLiu , Thank you for the reply. I have the same format as your dataset. But this format doesn't like zju_mocap. How do you load these data into training steps? Thanks!

[XingFen-Xie]

Hi @JiatengLiu , I have the same dataset format and write peoplesnapshot's data loader with reference to readCamerasZJUMoCapRefine, but I have the problem in the picture can you share your dataloader? Thanks!

[JiatengLiu]

Could you provide more info like the GT image? @XingFen-Xie I remember that if you code the dataloader in the format of the zju dataset you might get an upside down person, probably because of a different coordinate system. I could provide my script for reading snapshot dataset and how can I contact you? @XingFen-Xie @Yi-Cheng0101

[XingFen-Xie]

hi @JiatengLiu, I can not find the gt, but the people should be in the middle of the picture, not in the bottom. Could you please share your code with me via email? Thank you very much! My email address is 2983150767@qq.com. Thank you very much!