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!
May I ask for more details regarding 2D backbone pretraining? Quoted from the paper:
The 2D backbone for Human3.6M was pretrained on the COCO dataset [10] and finetuned jointly on MPII and Human3.6M for 10 epochs using the Adam optimizer with 10^−4 learning rate.
I have the following questions:
What loss function do you use in pretraining?
The setting "for 10 epochs using the Adam optimizer with 10^−4 learning rate" seems only to apply to the fine-tuning on MPII and Human3.6M, then what's the setting for the COCO dataset?
Can you provide the pretraining code or recommend a GitHub repository for pretraining?
I found that pretraining is very important. I trained the volumetric model without loading pretrained weights of the 2D backbone, and the relative MPJPE increase drastically from ~20 to ~30. Thus, 2D backbone pertaining is quite necessary.
Hello,
May I ask for more details regarding 2D backbone pretraining? Quoted from the paper:
I have the following questions:
I found that pretraining is very important. I trained the volumetric model without loading pretrained weights of the 2D backbone, and the relative MPJPE increase drastically from ~20 to ~30. Thus, 2D backbone pertaining is quite necessary.
Look forward to hearing from you. Thank you!