search

MomoAILab / ultrapose

Official repository for the ICCV 2021 paper: UltraPose: Synthesizing Dense Pose with 1 Billion Points by Human-body Decoupling 3D Model.
MIT License
103 stars 12 forks source link

annotations #4

Open liuzhihui2046 opened 2 years ago

liuzhihui2046 commented 2 years ago

Thans for your excellent work! According to the paper, the pose data were rendered according to the 3d model, and there is no error. However, when I visual the test image,I found that there were obvious errors。so I want to know what went wrong? "Third, the generated data annotations are absolute truth values without any error, and perfectly repre- sent the corresponding relationship between the 2D image and the surface of human body."

image

MomoAILab commented 2 years ago

Thanks for your question. That is mainly because we generate UltraPose using DeepDaz model. There is no annotation error by far. However, we register DeepDaz model into SMPL model format before we release the dataset. So there might be some mismatch between the annotations and images due to the registration loss. Especially when the body pose is sophisticated.

liuzhihui2046 commented 2 years ago

@MomoAILab Thank you for your reply！！！