mks0601 / NeuralAnnot_RELEASE

3D Pseudo-GTs of "NeuralAnnot: Neural Annotator for 3D Human Mesh Training Sets", CVPRW 2022 Oral.
168 stars 11 forks source link

Demo for InterHand2.6M #7

Closed redorangeyellowy closed 1 year ago

redorangeyellowy commented 1 year ago

Hello.

Is there any demo code for InterHand2.6M dataset?

mks0601 commented 1 year ago

which demo?

redorangeyellowy commented 1 year ago

In your code, there is a demo code for another dataset (i.e. COCO, Human3.6M, MPII, etc), but I don't think there is a demo code for InterHand2.6M dataset.

mks0601 commented 1 year ago

See this https://github.com/facebookresearch/InterHand2.6M/blob/main/tool/MANO_render/render.py

redorangeyellowy commented 1 year ago

Looking closely at the code on the link, it seems to render images with mano parameters.

But, if there is an image without a mano parameter, can I reconstruct the mesh of hand image?

mks0601 commented 1 year ago

All the demo codes in this repo is to render meshes. The above code is also for rendering meshes.

redorangeyellowy commented 1 year ago

https://github.com/facebookresearch/InterHand2.6M/blob/main/tool/MANO_render/render.py#L60

Then, if there is no value in the mano_params of the code above, isn't there an image that can't be rendered?

I wonder what I should do if I want to render these images which have no mano_params.

mks0601 commented 1 year ago

There could be some cases without mano parameters as there is no enough evidence (target) to get mano parameters. Such case mainly includes invalid 3D root joint coordinates as root joint coordinate is necessary to fit MANO parameters to GT 3D joint coordinates after canceling root traslation. In such case, you can fit MANO parameters to 2D joint coordinates, but the result would be not very good to call it as 'GT MANO parameters' as fitted data to 2D suffers from depth/scale ambiguity.

redorangeyellowy commented 1 year ago

Okay. Thank you for your reply.