The training result is weird

[j71200]

I've tried to train the model:

python create_avatar.py --subject {SUBJECT} --gender {GENDER}

Then, I get these files after training.

.
├── metrics.txt
├── OPTIMIZE_GAUSSIANS_0.ckpt
├── OPTIMIZE_GAUSSIANS_0.ply
├── OPTIMIZE_GAUSSIANS_3000.ckpt
├── OPTIMIZE_GAUSSIANS_3000.ply
└── state_dict_3000.json

After I import the OPTIMIZE_GAUSSIANS_3000.ply into GSAC Unity project, it is a large ball (see the image below).

If I set "Splat Scale" to 0.001, it becomes a lots of small points.

How to solve it? Any suggestion?

[hanhanhanhanhan929]

@j71200 你好 我也遇到了这个问题，请问您解决了嘛？我怀疑是否是数据的问题呢？因为我看生成的ply文件只有不到10m的样子

[j71200]

@hanhanhanhanhan929 您好 這個問題還沒解決（我的 ply 檔案也是不到 10MB）

[RendongZhang]

Thank you for reporting this issue! I updated a helper script (GSAC/helper_scripts/make_ply_after_training.py) for making PLY files for Unity platform(missing before) . Check it out!

To use it:

1. change : npz_file = "SMPLX_MALE.npz" to your avatar gender.
2. file_path = "state_dict_t36.json" to the state_dict.json file from your training. (it is saved under ./logs/subject/time/checkpoints/)

Make sure to put your trained state_dict.json and SMPLX_MALE.npz under the same folder of this script.

[j71200]

How can I get "temp.obj" and "extracted_texture.png" in helper_scripts/make_ply_after_training.py?

[RendongZhang]

Uploaded in helper_scripts folder.

[RendongZhang]

You can run it by just placing your trained state_dict.json. The extracted_texture is not actually used, and temp.obj is just a template mesh.

[j71200]

Thank you for your prompt reply. However, the result I obtained using the script still seems to have some issues.

[RendongZhang]

Hi there! I never had this issue before. If you don't mind, would you send me your state_dict.json file so that I can debug it closely? My email is rendong.zhang@vanderbilt.edu. Thank you.