search

mks0601 / I2L-MeshNet_RELEASE

Official PyTorch implementation of "I2L-MeshNet: Image-to-Lixel Prediction Network for Accurate 3D Human Pose and Mesh Estimation from a Single RGB Image", ECCV 2020
MIT License
724 stars 127 forks source link

Doubt regarding annotations #54

Closed Shubhendu-Jena closed 3 years ago

Shubhendu-Jena commented 3 years ago

Hi,

Thank you for the good work. I have a small doubt regarding the ground truth annotations. What is the difference between targets['orig_joint_img'] and targets['orig_joint_cam']? You're using targets['orig_joint_cam'] for supervision during the parameter regression stage of the model. Why could targets['orig_joint_img'] not be used instead? Looking forward to your response.

Thanks in advance

mks0601 commented 3 years ago

_img: (x,y) in pixel space, z: root-relative discretized depth space _cam: (x,y,z) in root-relative 3D coordinates

Shubhendu-Jena commented 3 years ago

Thank you for the response. I understood it. You have also mentioned this in your paper in section 3.3 Final 3D human pose and mesh. However, I am finding it a bit difficult to understand why the lixel stage can't be directly supervised using _cam: (x,y,z) in root-relative 3D coordinates. I'd be grateful if you could help me understand.

Also, I noticed that for the param stage, root-relative 3D coordinates have already been regressed by the following - root_joint_cam = joint_coord_cam[:,self.root_joint_idx,None,:] mesh_coord_cam = mesh_coord_cam - root_joint_cam joint_coord_cam = joint_coord_cam - root_joint_cam

However, in the evaluate function, the following has been done again - pose_coord_out_h36m = pose_coord_out_h36m - pose_coord_out_h36m[self.h36m_root_joint_idx,None] # root-relative

Isn't this redundant? If not, could you please tell me why it has been done?

mks0601 commented 3 years ago

Q1. Why lixel stage can't be directly supervised using _cam? A1. My network predicts lixel-based 1D heatmaps in a fully convolutional way. The 1D heatmaps of x- and y-axis are defined in image space.

Q2. Different joint set during the evaluation A2. I followed evaluation protocols of previous works. They used h36m joint set for the evaluation.

Shubhendu-Jena commented 3 years ago

Ah, alright got it. Thank you so much for the answers!