search

karfly / learnable-triangulation-pytorch

This repository is an official PyTorch implementation of the paper "Learnable Triangulation of Human Pose" (ICCV 2019, oral). Proposed method archives state-of-the-art results in multi-view 3D human pose estimation!
MIT License
1.09k stars 181 forks source link

about dataloader #136

Closed asw91666 closed 3 years ago

asw91666 commented 3 years ago

Hi! thank you for amazing contribution. I wonder why do you use islice() during training.

https://github.com/karfly/learnable-triangulation-pytorch/blob/9d1a26ea893a513bdff55f30ecbfd2ca8217bf5d/train.py#L174

thank you

shrubb commented 3 years ago

Hi, thanks. I believe it was used to limit the number of iterations per epoch. For example, if your metric is "accuracy after 20 epochs" and if you have two datasets where one of them is twice bigger than the other, that metric will be inconsistent. Because on a larger dataset, 20 epochs will mean twice more training steps.

asw91666 commented 3 years ago

then didn't you use all frame of data/human36m/processed files when you train ?

shrubb commented 3 years ago

Sometimes yes, sometimes not (like in experiments "without damaged actions"). Also, we used the CMU Panoptic dataset.

asw91666 commented 3 years ago

Do you mean you use H36m and CMU Panoptic jointly for learning, and testing on h36m?

shrubb commented 3 years ago

you use H36m and CMU Panoptic jointly for learning

As far as I remember, no. You should check the paper, we've explained everything there in detail.

asw91666 commented 3 years ago

OK Thank you very much !