Annotations / confidence

[isarandi]

Thanks for the dataset again.
When visualizing the data, I noticed that many labels are incorrect, probably the triangulation or camera calibration failed. Is there any way to obtain some confidence values to filter these out?

Example on 20220618_1b492fd601_subj22/c02:

Thanks!

[wangjiongw]

Thank you very much for your interest in our work, and we apologize for inevitable mistakes when publishing data.
After checking the local data, I think each 2D key point contains three numbers, corresponding to 2 ordinates and confidence. You can confirm whether it meets your needs.

In addition, what is currently released is the first version of the data, which contains some errors. For sessions with incorrect camera parameters, we roughly checked the data and recently uploaded a new file ignore_list.txt. We recently designed a pipeline to optimize camera parameters and keypoint annotation to optimize annotation. After the data is sorted, we will update the new version of the annotation as soon as possible, so you can stay tuned.

[isarandi]

Thanks for the reply, I'm looking forward to later updates. There aren't many large-scale in-the-wild datasets like this available, so it's quite exciting!

[isarandi]

@wangjiongw After some more checking, it seems the camera calibrations have errors. In some cases the horizontal and vertical focal lengths are specified as very different values which does not make sense for these images where the pixels are very nearly square shaped. It could be something with processing the checkerboard perhaps.