search

waymo-research / waymo-open-dataset

Waymo Open Dataset
https://www.waymo.com/open
Other
2.6k stars 605 forks source link

Evaluation on 3D keypoints. #513

Open wangjingbo1219 opened 2 years ago

wangjingbo1219 commented 2 years ago

Hi waymo research,

I have some questions regarding the keypoint data published with the latest releases (v1.3.2).

1) In your paper (https://arxiv.org/pdf/2112.12141.pdf), do you align the root joint between prediction and gt for evaluating OKS and MPJPE? If they are aligned, which joint should I use as the root joint? 2) In the metric API, the OKS considers both joint and box similarity. Does it mean this paper considers the box with and without keypoint annotation? 3) How to evaluate the results from the image-based method. For example, I use a pre-trained model to capture the SMPL of each human. This SMPL is built upon the space of the camera and I need to align this prediction to the gt 3d keypoints in Lidar space. Should I consider the rolling shutter for this projection?

Very thanks for your help. Best regards, Jingbo

birdflyto commented 1 year ago

hi, i have the same quetion about 1, the oks@2d in cocoeval need the statistics sigma for every joint point, do you have calculate the sigma and the oks@3d is just replace 2d distance and square with 3d distance and volume? Thanks!