search

zhangboshen / A2J

Code for paper "A2J: Anchor-to-Joint Regression Network for 3D Articulated Pose Estimation from a Single Depth Image". ICCV2019
MIT License
287 stars 46 forks source link

offset & code #32

Open shangchengPKU opened 4 years ago

shangchengPKU commented 4 years ago

Thank you very much for your excellent code, but I would like to ask you a few questions:

  1. What is the purpose of introducing offset?
  2. In the code file, you have center point, boundary box, mean / standard deviation What are the specific functions of GT key point documents?
zhangboshen commented 4 years ago

@shangchengPKU Hi, 1) The idea of predicting offset instead of pixel-wise probability is wide-applied recently (e.g., "Dense 3D Regression for Hand Pose Estimation", "AWR: Adaptive Weighting Regression for 3D Hand Pose Estimation", "Point-to-Point Regression PointNet for 3D Hand Pose Estimation"). In my opinion, predicting offset makes the output joints more accurate. 2) Sorry I don't understand your question very clearly. We use center point and bndbox to crop the hand/human-centered sub-image, mean/std normalize the original data.

shangchengPKU commented 4 years ago

@zhangboshen I understand the first question, the second question I read the relevant code, also understand, thank you very much, and I love you this article! Thank you again!

zhangboshen commented 4 years ago

@shangchengPKU Glad to hear that. And thank you for your kind attention to our work.