> There are duplicate labels in the training set, are there duplicate labels in the test set?

Hello,
Here are several examples where objects are labeled twice in the training set and we must clean them:

12863_d7ced_57b8f23416ef4.jpg
426947_dd8a3_592f44d99524a.jpg
416199_2ddf1_592f08c92167f.jpg
8475_e0a45_5774864a4f5c2.jpg
402604_81610_592dcbec5d78d.jpg
12863_3b530_57b8f1cc330cc.jpg
427178_17988_592f996bbe4bf.jpg
427178_d7ced_592f9972394eb.jpg

Has this problem been corrected for the test set? If not, a good submission might be prevented from getting the score it deserves.

Posted by: altvali @ June 19, 2018, 5:52 a.m.

Hello

Thank you for your observation ! We are aware of some duplicates in the train set. From our evaluation, there are about 250 duplicate sign markings from a total of more than 50k markings on the train set. What we can confirm is that there are no duplications on the test set labels.

If you have any more questions or observations, please don't hesitate to post on the forums !

Posted by: telenav @ June 19, 2018, 2:51 p.m.
Post in this thread