No contact

[JunxingHu]

No contact

[xiexh20]

Hi,

Thanks for your interest!
I tried the demo several times, but I've never seen this before :),
have you tried rerun the demo?

[JunxingHu]

Thank you for your prompt response! I tried the demo on two servers and several samples from the BEHAVE dataset as below, but none of the blue areas representing contacts were reconstructed. I only modified --obj_name in demo.py for different samples. The inputs include k1.color.jpg, k1.color.json, k1.mocap, k1.mocap.ply, k1.obj_mask.jpg, and k1.person_mask.jpg. Do I need to set other items? Thank you for your kind help.

[xiexh20]

Hi, I think your command is correct, as most of the examples you show are reconstructed reasonably.
what do you mean blue areas representing contacts? Do you refer to the contact sphere shown in the teaser figure? It was added for better visualization, you can refer to here to check how it is generated.

[JunxingHu]

Thank you for your help! I will try adding the contact sphere. However, the reconstructed human body does not touch the basketball as shown in the first question. Is there any other version of the pre-trained model? In addition, since I use the RTX3090, the version of some algorithm packages such as cuda cannot be completely consistent with the requirements. Have you tested it in other environments such as cuda11.0? Thank you very much.

[xiexh20]

Hi, I tried this pre-trained model on cuda11.3, Nvidia RTX 8000 and V100. I don't have models trained on other machines. But I tried to run this several times and always got the same result.
Maybe you can try to rerun this example again and see if you get different results. Or if you can reproduce the numeric results on the BEHAVE test set, it should be fine.

[Young2647]

I tried the basketball example and get the same result as @JunxingHu. No contact on the basketball

.

[JunxingHu]

Hello, thank you for your patient help. Since most GPUs do not support environments lower than cuda11.0, could you please provide the versions of all packages in anaconda when cuda11.3 to help us reinstall the environment? Moreover, testing 90 sequences is very time-consuming, so I ran one sequence BEHAVE/sequences/Date03_Sub03_backpack_back/ by using python recon/recon_fit_behave.py chore-release --save_name chore-release -s [path to one sequence] and get the results by using python recon/evaluate.py as below. Are these results the same as yours? Thank you.

{'smpl': {'mean': 0.06429596999150605, 'std': 0.021513825524556374}, 'obj': {'mean': 0.07156142227877363, 'std': 0.03903869274395644}, 'total': 41}
{'smpl': {'mean': 0.06429596999150605, 'std': 0.021513825524556374}, 'obj': {'mean': 0.07156142227877363, 'std': 0.03903869274395644}, 'total': 41}

[xiexh20]

Hi, this is the result I have for the sequence you mentioned:

{'smpl': {'mean': 0.06414959380968863, 'std': 0.02062545638459632}, 'obj': {'mean': 0.07133498928598118, 'std': 0.040964391942100074}, 'total': 41}
{'smpl': {'mean': 0.06414959380968863, 'std': 0.02062545638459632}, 'obj': {'mean': 0.07133498928598118, 'std': 0.040964391942100074}, 'total': 41}

they are pretty much the same as your results, so I think your environment is fine. It might have some difference in one or two specific examples, but overall results will be the same.

[JunxingHu]

Wow, thanks for your prompt reply. This is a great work!

[luckcodingdog]

I also tried the basketball example and get the same result as @JunxingHu & @Young2647 . No contact on the basketball.
I run the demo on 3090 GPU with pytorch 1.11 and CUDA 11.2.
At the same time, i do the same evaluation with JunxingHu, my result is similar with yours.