How is the accurate?

[tranducanhbk]

Hi mks0601, I have jsut finish train Hand4Whold follow your instruction with dataset H36M+MPII+MSCOCO but when i run demo the results is not good. U can see images below. I trained 6 epochs eachs state

[mks0601]

Did you set bbox in the demo code correctly?

[tranducanhbk]

Could u tell me detail about this parameter and how to set it correctly?

[tranducanhbk]

after i set bbox corectly the hand is good but leg and feet is wrong

[mks0601]

I think the result is worse than mine. Could you test that image with the pre-trained model?

[mks0601]

Or you can test the demo image with your model and compare the result with mine. I remember the result of the above image was good

[tranducanhbk]

I compared my model and your model and the result is below. this is mine: and this is yours (i got your model from this link Download Hand4Whole trained on H36M+MPII+MSCOCO from here.) the result is same

[mks0601]

Oh I see. I might be confused with my other models with better performances. I guess the wrong result of the right leg is due to the occlusion.

[tranducanhbk]

Can u tell me which is better models u have?

[mks0601]

That will be submitted to another conference soon. The codes and pre-trained models of that paper will be released upon acceptance. Please stay tuned :)

[tranducanhbk]

i am looking forward to your updating thank you very much

[tranducanhbk]

Hi mks0601! When i test my model trained with H36M+MPII+MSCOCO dataset but the result i not good so i would like to make custom dataset to train. Could u give detail guide for make custom data to train this is result

[mks0601]

Please prepare your custom dataset in a MSCOCO format (http://cocodataset.org/#format-data). And make a dataloader like https://github.com/mks0601/Hand4Whole_RELEASE/blob/main/data/MSCOCO/MSCOCO.py

[tranducanhbk]

thank you for your quickly reply

[tranducanhbk]

Hi mks0601! I have just make a custom dataset follow COCO format (just only label for boby part) then I train HAND4WHOLE with H36M+MPII+MY_MSCOCO dataset but when i run demo the error occur like images below Do u have any ideal for this error? I have to label face, hand and foot part or not?

[mks0601]

I guess you should check the dimension of the mesh. Without face/hand/foot labels, the output mesh will have weird face/hand/foot

[tranducanhbk]

I checked mesh output all element is nan Some thing wrong when we train with COOC data withoit face,hand and foot labeling?

[tranducanhbk]

this is my custom json file for trainning coco_wholebody_train_v1.0.txt and about MSCOCO_train_SMPLX_all_NeuralAnnot.json file, i have to make a new file for new custom data or not if u found any abnormal please tell me

[tranducanhbk]

Do u have any update for this issue? I am looking forwarding your reply! Thank you

[mks0601]

Sorry I have no time to check the details of your custom dataset. MSCOCO should have body/hand/foot/face annotations. If you want to get the body-only model, check here

[tranducanhbk]

maybe u misunderstand my issue, I would like to made HAND4WHOLE version but when I test some image in my data with weight trained public dataset (H3.6,MSCOO,MPII) the result is not good. So I would like to make custom data from my data to train. I think I got mistakes when making custom dataset. In your dataset have MSCOCO_train_SMPLX_all_NeuralAnnot.json but in m data don't have. if u have time, please tell me how to create MSCOCO_train_SMPLX_all_NeuralAnnot.json for custom dataset Many Thanks

[mks0601]

without SMPLX parameters, the model will perform really bad. They are 3D pseudo-GTs of MSCOCO dataset. My another paper (NeuralAnnot) is used to obtain the 3D pseudo-GTs.

[tranducanhbk]

thank you so much