Loss function

[alebas-fr]

Hello,

I don't understand well your loss function, Can you give me some details please! How you calculate the loss!

Another question: if I don't use the graph in my architecture, is it possible to use this same loss formula that you use please!

Thanks a lot!

[abduallahmohamed]

Thanks for asking:
Section 4 in the paper explains the loss function.
A TL;DR:
The loss function have 3 components:
1- MSE between the joints and the prediction
2- A consistency loss that:
2-a Make sure the intradistance between joints are consistent
2-b Make sure that the angles between the joints are consistent
AKa, we don't want to end up with weird angles and/or weird skeleton bone lengths. Hope this explains it.
Foe the second question, the loss function is independent from graph configuration. So, you can use it on any other model.

[alebas-fr]

Thank you, but in your code, there is 2 cases : if if args.use_cons or not;
In this case in your code you use juste the loss in the part of this part of else:
l2loss = nn.MSELoss().cuda()
def graph_loss(V_pred,V_trgt):
return l2loss(V_pred,V_trgt)

and this because the use cons argument is set to false in your code !

Can you give some explanation please!

thanks

[alebas-fr]

What is the formula you use to calculate the loss and get the results you put in your article please ( there is many cases in your code in the loss function)

[abduallahmohamed]

args.use_cons is to enable to disable the consistency loss.
This is for ablation study, see Table 3 for the ablation study and the effect of different parts of the consistent loss.
The results in tables 1& 2 are the best result from Table 3.
So, for each dataset you can get the flags for enable/disable the cons. loss or the internal parts of it (cosine, L2) based on the ablation table.
For the GTA-IM, the best results are obtained when the cons. loss and all of it is parts are enabled.

Does this help?

[alebas-fr]

Yes thank you. On the other hand for the loss of PROX I saw that only L2 leads to good performance (NOT LIKE gta) but I didn't understand why! can you explain to me please!

Also, I have another question is your visualization script can be used on the PROX basis please, I want to do the visualization on PROX rather how I have to do it please

[abduallahmohamed]

I didn't invest in visualizing PROX.
For the PROX performance, quoting from the paper:

AKA, it's a noisy and biased dataset unlike GTA-IM.