ChaLearn Looking at People: Identity-preserved Human Detection (Track 2: Thermal) Forum

Go back to competition Back to thread list Post in this thread

> Wrong class labels in dataset

Dataset has only one class - human. So the class label should be 0 for every bounding box. But there are over 10K bounding boxes which have 1 as class label.
An example is "iphd_train_thermal-v2\labels\vid00083_0xqafLPk.txt"

Posted by: ashish.23ks @ Jan. 3, 2020, 11:14 a.m.

(Sorry for the delayed response. I have been out of office since Jan 1st.)

You are completely right. However, class=1 annotations can be safely replaced with 0.

The explanation for these is that we were planning to release extra annotations for the dataset. In particular, 1-class labels are of subjects lying on the floor after a fall situation.

Thanks for noticing and making the point.

Kind regards.

Posted by: aclapes @ Jan. 6, 2020, 11:41 p.m.
Post in this thread