Hi, I recently found another bug in the dataset.
VID: 59323787d5
Segmentation File: 00070.png
It has pixels labeled as '10', but it does not show up on the meta info. I believe it should be just 1 rather than '10'.
Thanks.
Posted by: hi.nine @ Aug. 9, 2018, 2:21 a.m.For a few videos, there exists a label which does not show up on the meta information. I guess it happened when you filtered out small objects to build the dataset. Filtering out such an annotation can be done easily, so it should not be a big deal.
However, I believe that some of the cases look like just a noise (data annotation error) like the example above. Here is the list I found for now.
VID, Frame, result of np.unique('***.png')
59323787d5, 70, {0, 1, 10}, 10 is a noise
909dbd1b76, 50, {0, 1, 12}, 12 is a noise
909dbd1b76, 60, {0, 1, 3}, 3 is a noise
afe1a35c1e, 65, {0, 1, 2, 3, 11}, 11 is a noise
bb1c770fe7, 145 {0, 1, 2, 3, 5, 9}, 9 is a noise
b2edc76bd2, 95, {0, 5, 6, 7, 9, 10}, Only 0 and 5 seems correct
Hi,
Thanks a lot for your careful check! For this dataset, we have some noisy annotations due to accidental clicks by labelers. However, these clicks cause only few pixels' error on the annotations map. Actually part of the reason we remove the small objects in meta data is that we can remove most of these errors too. So when using this dataset, please only use objects shown in the meta data, and consider all other pixel values as background. Sorry for the inconvenience caused. We will try to remove such kind of errors in the annotations in the next version. Thanks!
Best regards,
Organizers of the challenge