There seems to be a bug in the data. The structure of files predictions_train.json and predictions_val.json should be:
“VC_a conf_a x_0 y_0 x_1 y_1”,
“VC_b conf_b x_0 y_0 x_1 y_1”,
but in the data, there is missing whitespace between x_1 and y_1
Hey @lsoukup, sit tight!
I’ll let the team know about your issue incase they didn’t see this one.
Additionally, I’ve also found out that y_0 and x_1 are always equal, which seems a little suspicious to me. Are you sure, that the training and validation data are correct?
First of all, sorry for the very late reply but the notifications of the discussion section were deactivated!
We have now updated the dataset and corrected the two issues related to the predictions files.
Note that the issues were problematic only if you try to use the size of the detected objects in your runs.
We tried to use the updated data version, unfortunately, the gt users ids do not match anymore with the users ids in the prediction files.
Thanks for noticing the problem with user IDs, it is now corrected.
Please download the dataset again.