problem? loading ICVL test_seq dataset about deep-prior-pp HOT 11 CLOSED

Of course you can run the second file as well, there is no difference, it is another test sequence, however very similar.

from deep-prior-pp.

hello, @ @moberweger
now I met something confusing after reading the code,
in the file main_icvl_com_refine.py row147
poseNetParams = ScaleNetParams(type=1, nChan=nChannels, wIn=imgSizeW, hIn=imgSizeH, batchSize=batchSize, resizeFactor=2, numJoints=1, nDims=3)
why you set numJoints = 1 rather than 16?

from deep-prior-pp.

whether it means your pretrained model is trained on the only first joint coordinate?
In my opinion I think this is not quite proper. @moberweger

from deep-prior-pp.

@WeihongM L147 that you mentioned is only for the refinement of the hand localization (com). Therefore, we only predict a single joint, ie the hand center. Thus numJoints=1 and nDims=3 for the 3D offset.

from deep-prior-pp.

@moberweger sorry, so which file is the code to predict the all numjoints? or is it right to just change the numjoints = 16 to train the network, if not, can you give me some advice how to write?
and Now I am more confused, in this repo, you use the multi-scale input to train the offset of a single joint, but in the paper deep prior++, the offset is not trained in this way.
what is more, I find this code(L147) is also in the deep prior repo code, you also use numjoint=1 and multi-scale input to train.
I guess you just want to give the first joint to check the loss and error in this code?
Can you understand what I mean, hopefully look for solutions, thx

from deep-prior-pp.

You should take a look at main_icvl_posereg_embedding.py if you want to predict all joints.
Regarding the multi-scale input, it actually does not matter in terms of accuracy. It is the same code from DeepPrior. If you read the paper, you can see the usage of the different networks.

from deep-prior-pp.

ok, @moberweger
Now I try changing the code (change the numJoints from 1 to 16) to check the performance, because in the paper deep prior, you have mentioned this network to predict all the joint coordinate.
Do you mean the multi-scale input does not have much improvement on the joint accuracy?

from deep-prior-pp.

The script main_icvl_posereg_embedding.py does already what you want, I guess. Also, multi-scale does not help much, yes.

from deep-prior-pp.

@moberweger hello,
when I run the main_icvl_posereg_embedding.py on the NYU dataset, I got the mean error 13.8069019318mm, but in the paper deep prior ++ which is 12.3 mm on the NYU dataset. Can you give me some advice on improving?

from deep-prior-pp.

main_icvl_posereg_embedding.py is intended for ICVL dataset. So please check that. Did you use the refinement network for the hand localization?

from deep-prior-pp.

thanks for your reply very, i know what i missed

from deep-prior-pp.