Code release?

[Arthur151]

Hi, congratulation on the great work in 3DV! May I ask when are you going to release the code? Looking forward to learn from it. Best wishes, Yu

[nicolasugrinovic]

Thanks for your interest in our work. Our code will be out in a few weeks.

[Arthur151]

Thanks! Looking forward to the code release!

[Arthur151]

@nicolasugrinovic Could you please provide the used frames / depth ground truth of 3DPW and MuPoTS-3D that you used for evaluation? And the function that you used to compute the d_{ord}. So that we can compare with the results reported in Tab. 1. At least the used frames / depth ground truth of 3DPW and the function, which is supposed to be just a few line of code, I suppose? Really looking forward to this!

[nicolasugrinovic]

Hi,

For MuPoTS-3D we used all frames where the matching between estimated subjects and GT subjects was possible. Meaning that when estimating each persons pose/shape, we do not know which GT each corresponds to. For doing this, we compute distance of the 3D joints and the 2D bbox overlap of each estimated person w.r.t. to GT labels. It is hard to do that matching when people overlap in the image, so we discard frames where this happens.

For 3DPW we use only the frames where there is GT for 2 persons in the test set.

Depth ground truth is taken from the 3D joint annotations, we take the root depth as the person's depth. Once we have these correspondences, we score the depth order in a pairwise manner.

We will soon release the code.

[Arthur151]

Thanks for your quick reply! It helps. Could you provide the function (code) that you score the depth order in a pairwise manner, like given the Depth ground truth and the predicted depth value?

[Arthur151]

@nicolasugrinovic Sorry for bothering. Is there any change that the code would be released in next week? At least the code describing evaluation matrix, which is supposed to be quite easy to organize?

[nicolasugrinovic]

I'll try to get it ready for this week. Sorry about that.

[Arthur151]

Thanks! Really looking forward to it.

[ssssholmes]

@nicolasugrinovic Thanks for your great work. Will the code be released this week?

[Arthur151]

WOW, it's coming. Looking forward to the evaluation part! Thanks in advance, Doc!

[Arthur151]

Dear Doc. Nicolás Ugrinovic, Thanks a lot for your efforts! Best, Yu

[nicolasugrinovic]

Hi @Arthur151, sorry for the delay. I'll try to upload the d_{ord} eval function as soon as possible and release the rest after that.

[nicolasugrinovic]

@Arthur151 you now have the function to calculate depth order here. Basically a matrix operation to know which persons are behind or in front of the other persons.

[Arthur151]

Thanks for the code! It seems that the depth order only describes the closer or farther, without equal class. The if ... continue in code makes it hard to determine which images are used for evaluation.

[nicolasugrinovic]

What do you mean by equal class?

Yes, some images where it is not possible to do an ID matching of the estimated person with its corresponding GT are not considered. This happens usually when a person is occluded by another person. This is similar to the matching used in the code for 3DPCK metric. Once you generate data and run the eval, you can tell which images are taken into account. I could also make a list of the images being evaluated, throughout the week.

[Arthur151]

When we try to determine the ordinal depth between each pair of subjects, we may need to classify the subjects to equal class that their difference in depth is less than a pre-defined threshold.

Thanks for your help!

[nicolasugrinovic]

@Arthur151 you can check the file done_files_eval.csv. It is a list of all the images used for the evaluation. The evaluation code discards some examples in which it is impossible to assign a corresponding GT to a person due to heavy occlusions. Hope this helps

[Arthur151]

Thanks for your efforts! Yes, it would be helpful.