apple / ml-neuman

Official repository of NeuMan: Neural Human Radiance Field from a Single Video (ECCV 2022)
Other
1.26k stars 141 forks source link

How to get Da-pose canonical human from a single image? And use the canonical human to do what? #31

Open cocoshe opened 2 years ago

cocoshe commented 2 years ago

Use the key points and bounder of human in the image, and compare it with SMPL mesh to train the parameters in SMPL for a suitable “meta human”?Then control the parameters to edit the pose of the "meta human"? I don't know if my view’s right.

jiangwei221 commented 2 years ago

I’m not sure what do you mean…

cocoshe commented 2 years ago

I’m not sure what do you mean…

I'm wondering how to get the canonical human

jiangwei221 commented 2 years ago

我还是不明白您的问题是什么意思?

cocoshe commented 1 year ago

我还是不明白您的问题是什么意思?

就是我想请教一下,如何用图像上2d的人物生成paper中“大”字形的3d人体模型?是通过SMPL拟合得到参数后,调节姿态参数得出来的吗?

jiangwei221 commented 1 year ago

哦哦,我们用ROMP预测人的beta,然后pose的话全是0,然后手动调节了胯部关节的角度: https://github.com/apple/ml-neuman/blob/84d4685074327227457276f1b60d467f3dab1211/data_io/neuman_helper.py#L292-L297

cocoshe commented 1 year ago

ROMP建立起人体的mesh模型,这个过程都是靠图像中的2d人物吧?我不太明白既然有了可编辑的mesh,大字型和T字型为什么效果会有所不同呢?感觉只是手动调整了可视化的结果,都是表示同一个mesh;顺便再问一问怎么让人体做出自己想要的动作呢?难道是人为控制pose的参数吗?还是说从其他的视频里面通过某种方式提取出动作信息,来控制这里的人体?

jiangwei221 commented 1 year ago

ROMP建立起人体的mesh模型,这个过程都是靠图像中的2d人物吧?

是的,ROMP只需要一张图就可以预测SMPL参数。

我不太明白既然有了可编辑的mesh,大字型和T字型为什么效果会有所不同呢?感觉只是手动调整了可视化的结果,都是表示同一个mesh;

大pose和T-pose的区别我们在paper和github issues中都有阐明:#21

顺便再问一问怎么让人体做出自己想要的动作呢?难道是人为控制pose的参数吗?还是说从其他的视频里面通过某种方式提取出动作信息,来控制这里的人体?

只要能提供一组新的pose,就可以渲染对应pose的人体。具体pose如何获得,可以是手动调节获得,也可以通过专业动捕设备,或者是ROMP。

cocoshe commented 1 year ago

SMPL的mesh可以带颜色吗?还是通过其他什么方式给mesh上色?

jiangwei221 commented 1 year ago

Emmmm,如果你有texture map是可以的,但是neuman不生成texture map,只生成nerf。

ihongxx commented 1 year ago

请问下,先用ROMP对图像进行估计,在利用SMPL处理及SMPL优化方法处理,我理解SMPL优化方法中只对pose点进行了优化,论文中也展现了pose点优化对比,但是模型的形态是否有进行优化处理呢?

jiangwei221 commented 1 year ago

我们没有优化beta,只优化了pose