search

mks0601 / 3DMPPE_POSENET_RELEASE

Official PyTorch implementation of "Camera Distance-aware Top-down Approach for 3D Multi-person Pose Estimation from a Single RGB Image", ICCV 2019
MIT License
824 stars 147 forks source link

MSCOCO keypoints with pre-trained model #68

Closed d-cota closed 4 years ago

d-cota commented 4 years ago

Hi, I'm trying to obtain the eyes key points from PoseNet and I've seen that there are annotations for that in the MSCOCO dataset. I've also noticed that the pretrained model has been trained using MuCo + MSCOCO so I'm trying to figure it out if I can obtain from this model the eyes coordinates. In the demo script appear only the MuCo joint set, there is away to obtain the MSCOCO body points or I need to re-train the net? Thanks you

mks0601 commented 4 years ago

Hi,

Unfortunately, MSCOCO is 2D pose dataset, so PoseNet cannot be trained on eye joint for the 3D pose estimation. Instead, you can check my new repo (https://github.com/mks0601/I2L-MeshNet_RELEASE). It predicts 3D human pose and mesh, and 3D eye coordinate can be obtained.

d-cota commented 4 years ago

Thanks you. But if I don't need the 3D estimation of the joints can I train the net only on the 2D dataset?

mks0601 commented 4 years ago

For the 2D estimation, yes you can but you should change code a little bit. I think hrnet would be better for you if you are interested in only 2D estimation.

d-cota commented 4 years ago

Thanks you, appreciate so much your work!