possibility of extracting fl-curves

[Zevrap-81]

Dear Authors,

I want to just predict the fl-curves that optimized for a given image. Is it possible from your code to only extract 3d curves in the posed space? If yes could you give me some hints on how to do so?

p.s. I have used align_fl function to get the curves but they seem to be in canonical pose

[lingtengqiu]

If u wanna get 3D curves in observation space, you can use Deformation Field which consists of both non-rigid deformation and skinning weights. For skinning weight, u need to predict human pose(SMPL theta). In terms of non-rigid deformation, u can employ average non-rigid latent code.

[Zevrap-81]

Could you please some more details on the non-rigid deformation and how it can be accomplished ?

[lingtengqiu]

More details you can see our paper. Actually, we initialize feature embedding that is trainable parameters for non-rigid deformation learning. Given image embedding condition on MLP to predict non-rigid deformation.

[msverma101]

can you please check this post. here i tried the feature lines using mmpose and also hrnet-agg-81 both of which doesnt seem to cut it for fl points.
i dont understand why your code is so convoluted in terms of getting simple fl lines and curves.

https://github.com/Mamba-ZJP/deepfashion2-kps-agg-81/issues/2

is there any other avaiable code that you can please clearly share here that finds these landmarks of the garment which you haven not mentioned how you got them as a json file. were they manually annotated or other pretrained model for it