search

gulvarol / bodynet

BodyNet: Volumetric Inference of 3D Human Body Shapes, ECCV 2018
http://www.di.ens.fr/willow/research/bodynet
MIT License
261 stars 42 forks source link

FV and SV image input #8

Closed bnathoo closed 5 years ago

bnathoo commented 5 years ago

Hi Gul

Is it possible to input front and side view images to generate the final product like you have done in training to provide more accurate volume representations. If so how do we do this?

Thanks

gulvarol commented 5 years ago

Hi, I am not sure I understand the question. 'FV/SV image as input and voxel as output' is something I didn't try. You can try by adapting the code. It is similar to 'segmentation in, voxel out' as in here. My use of FV/SV is more like a constraint on the output.

bnathoo commented 5 years ago

Thanks for the quick reply Gul. Just to clarify, when I refer to FV and SV image input I assumed you input 2 images for silhouette reprojection as noted in table 2 of your paper (lines 6, 8 & 10) and also in the training section of the readme file under subnet 5 and bodynet.

You mentioned your use of FV/SV is more like a constraint on the output which is great, but how do you input the FV/SV images together to act as the constraint when currently the system does not allow for multiple image input?

Hopefully that clears up my query a bit. Thanks

gulvarol commented 5 years ago

No, I do not input the FV/SV images to the network, they act as ground truth. So I add them to the ground truth table, e.g. here. I set the fields of the instance in the data loader.

bnathoo commented 5 years ago

Thanks Gul