Inaccurate fitting on images

[sparshgarg23]

Hi,I tried your code on some images obtained from web-cam as well as some images from yoga-82 dataset(on which mediapipe) was trained. Since,the pose changes are extreme,it seems the model is not able to get a correct fit. Any reaons on why that is happening results of fitting romp

original images

 ? I hope you can give me some insight into what can be done to improve the fitting ,should i tune the parameters in the config file?

[Arthur151]

Have you fine-tuned ROMP on yoga-82 dataset with 2D pose annotations? To get a more accurate fitting, we may need the assitance of 2D pose. Taking ROMP's output as initialization, iterative optimizing the results to fit the 2D pose annotations, which can be achieved via SMPLify

[sparshgarg23]

ok will look into it thanks,one last thing can I use the output of ROMP as an intialization for SMPL-X. Also,apart from SMPL-X,is there any other work that deals with capturing full body expressions that you're aware of

[blibot]

hello @sparshgarg23, how did you manage to make the mesh transparent ?

[sparshgarg23]

@blibot I am not sure I understand.The color of the mesh is the same as used in SMPL-X.Because I didn't fine tune ROMP,that's why the fit isn't accurate and it looks as if you just overlaid the mesh on the image.

[blibot]

in this first picture it looks like the mesh is a bit transparent. I was wondering if you changed the opacity in the code.

[sparshgarg23]

i didn't do any change.

[blibot]

ok thanks. are you running Simple_ROMP or ROMP ?

[sparshgarg23]

as of now i am not working on the project