How to train custom data & annotation without bbox?

[Jeong-Bin]

My project is to get x, y ,z keypoint-values of body, hands and foot from 2D videos. And also I'm trying to train by my custom data & annotaion. In this process, I have 2 questions.

First, Which model should I follow? 2D Human Whole-Body Pose Top-Down Video Demo? or 3D Human Pose Two-stage Estimation Video Demo?

Second question is about bounding box. My json annotation files don't have the 'bbox'... How to train customize dataset without bounding box? Do I just change my json file like COCO format without bbox? or just create and add custom_config.py like this tutorial?

I referred to this issue(#1473), but it is still difficult because I'm not used to this Human Pose task.😢

[Ben-Louis]

Hi, thanks for using MMPose. For your questions,

1. Since you want to estimate 3D human pose, you could try the later one, 3D Human Pose Two-stage Estimation Video Demo

2. Whether to use bounding box depends on the pose model you want to use. If you want to train a bottom-up pose estimator, bounding boxes are neglectable in your json file. If you want to train a top-down pose estimator, which is adopted by the above demo script, bounding boxes are necessary in this case. You may calculate the bounding box via keypoints as following: bbox=[x, y, w, h] where

   • x is the x-coordinate of the left-most visible keypoint
   • y is the y-coordinate of the top visible keypoint
   • w is the difference of x-coordinate between the right-most and the left-most visible keypoint
   • h is the difference of y-coordinate between the bottom and the top visible keypoint

[Jeong-Bin]

Thank you for your kind help! Then, if i change my annotation format to COCO, should I keep this format thoroughly? Are all attributes necessary?