2022 IEEE GRSS Data Fusion Contest Track SLM Forum

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

> annotations of labeled data

Dear organizer:
we find miniFrance seems to have 15 classes(without clouds shadows or no data), and our datasets only have 14, and we are also confused about where we should find the class name for each class. In addition, we notice that there are images labeled all with 0, is this a problem? Thank you for your time! Best regards!

Posted by: Einfall @ Jan. 10, 2022, 2:59 a.m.

Please check here: https://www.grss-ieee.org/community/technical-committees/2022-ieee-grss-data-fusion-contest/

Posted by: rhaensch @ Jan. 10, 2022, 7:23 a.m.

thank you very much! just for verfication, the images labeled all with 0 means this image is ignored?

Posted by: Einfall @ Jan. 10, 2022, 8:30 a.m.

It means that for this image (or this image part) no label information is available (similar to the unlabeled part of the training set). How this is handled, is up to the participant. Ignoring it is one option. However, the spirit of semi-supervised learning is to leverage this information, too.

Posted by: rhaensch @ Jan. 10, 2022, 8:36 a.m.

Dear organizer, when I submit my result, I keep getting the message "Image format should be tiff". I checked the image format and it is a Tiff file and can be read by pillow. I am wondering if you can provide me with an example of a submission. Best regards!

Posted by: Einfall @ Jan. 13, 2022, 12:32 p.m.

The submission file format should be the same as the reference data just with different folder and file names (and no class IDs of 0 and 15).

Posted by: rhaensch @ Jan. 13, 2022, 12:42 p.m.

for example, image 49-2013-0426-6736-LA93-0M50-E080.tif should be placed in submission>>Angers>>BDORTHO>>49-2013-0426-6736-LA93-0M50-E080_prediction.tif
Is that correct?

Posted by: Einfall @ Jan. 13, 2022, 12:46 p.m.

No, as your prediction is not from BDORTHO (i.e. a RGB file). It should be in <fname>.zip->Angers

ps: I noticed your submissions are an order of magnitude larger than the ones from anybody else. Please make sure to use compressed tifs.

Posted by: rhaensch @ Jan. 13, 2022, 1 p.m.

Dear organizer, I am wondering if we can use extra data for training?
best regards!

Posted by: Einfall @ Jan. 15, 2022, 11:54 a.m.

Yes, as long as you do not violate the semi-supervised framework, i.e. semantic labels only for the defined part of the training data, and you are open about it: Before the beginning of the test phase, you will need to submit a short description of your method where you need to state all dataset and how they have been used.

Posted by: rhaensch @ Jan. 24, 2022, 7:12 p.m.
Post in this thread