Model accuracy

[Ulduzpp]

Hi @jpcurbelo
Thanks for sharing your great work in github, it is splendid, I just wanted to ask you about the accuracy of model you provided, and also do we need to retrain the model or finetune it?

[jpcurbelo]

Hi @Ulduzpp. Thanks for your interest in our work! We did run some experiments to assess accuracy (see plot). You can find the details in our paper (also see the plot). In case you can not access the paper, please, share your email and I will send you a copy.

Also, see that the output file includes the relative percentage deviation (δ%). This is the measurements taken from the generated avatar compared with the predicted/imputed ones.

Please, let me know if you have further questions, Happy to help.

[jpcurbelo]

About retraining the model. The script to train the extractor model can be adapted and used for different configurations. You might want to play around to include different measurements to be used as inputs and labels as well.

[Ulduzpp]

Hi @jpcurbelo thanks for your comprehensive response and about the paper this is my email: [email protected]
I would like to have a copy of your work to read

[Ulduzpp]

also I mostly am curious about the accuracy of body measurements I saw that you have written it has 97% accuracy for body measurements but I would like to know if it consists children as well or not besides are there especial rules that we should consider while giving image inputs to the model? cause I tested it but the weight and chest girth were too different from reality that is why I ask that maybe it is because of my input images.

[jpcurbelo]

Hi @Ulduzpp. I just sent the paper by email.

About results and input files:

1- The NN to extract the measurements was trained on ANSUR II data that only contains adult subject. Therefore, I don't expect the model to perform well on children data. It would be interesting though to re-train the model on a children dataset to see what happens but I guess it won't be easy to find images available.

2- For the input images, please, see this comment. I would like to see the silhouettes extracted during from experiments if you don't mind to share the files - just the silhouettes, not the original photos of the person. Thanks.

[Ulduzpp]

Thank you @jpcurbelo for your email and the detailed comments, I have received the paper.
also about the my experiment this is the silhouettes I have got, besides the height is 171 and male
silhouettes.zip