Inquiry about the inaccurate 3d human pose labels about 4d-or HOT 2 CLOSED

Hi, thank you for your interest. The human poses are semi automatically annotated and might not always be perfect. In the bottom examples you sent I think the issue is the human is actually not visible in the point cloud, and the human pose track stays at the last spot where the human was detected. This might be an error in the code but also potentially an error in the annotations.

However in the other examples I think it is not as bad. You can try to visualize the pose from only one human to see the accuracy. I think overlaying all points from all humans makes it hard to judge.

Finally if you are looking for better human pose annotations, you should check out https://github.com/CAMMA-public/MVOR. The annotations are more precise, however the dataset is much smaller. So you can see which fits your needs the best, maybe both together?

hope these help

from 4d-or.

Hi, thank you for your interest. The human poses are semi automatically annotated and might not always be perfect. In the bottom examples you sent I think the issue is the human is actually not visible in the point cloud, and the human pose track stays at the last spot where the human was detected. This might be an error in the code but also potentially an error in the annotations.

However in the other examples I think it is not as bad. You can try to visualize the pose from only one human to see the accuracy. I think overlaying all points from all humans makes it hard to judge.

Finally if you are looking for better human pose annotations, you should check out https://github.com/CAMMA-public/MVOR. The annotations are more precise, however the dataset is much smaller. So you can see which fits your needs the best, maybe both together?

hope these help

Okay I get it. Thank you for your kind reply!

from 4d-or.