Loading
Cookies help us deliver our services. By using our services, you agree to our use of cookies. Learn more

crowdAI is shutting down - please read our blog post for more information

Mapping Challenge

Building Missing Maps with Machine Learning


Completed
719
Submissions
1071
Participants
44927
Views

solution journal [0.868]

Posted by neptune.ml over 1 year ago

Hi there,

We took a closer look at a few things and managed to improve the score significantly:

We started experimenting with using VGG and Resnet as the encoder part of the unet network We added distance-to-edge-weighted loss (current LB is the result on resnet34) We implemented the per pixel loss (still expeimenting) We added deeper encoder architectures Resnet101, Resnet152 (still experimenting) We added TTA up-down and left-right flips + rotation [0,90,180,270]

The best results we’ve gotten so far are trained without dice loss which results in not very smooth (though accurate) predictions. TTA smoothens then to a certain extend but we need some good models with dice to entirely get rid of that problem. Another option is better morphological post processing which we are also working on right now.

Everything that we do is as always available for you to use/change/comment in our open-solution repo https://github.com/minerva-ml/open-solution-mapping-challenge . Newest ideas are usually on the dev branch so make sure to check that out too.

Best, and good luck!

1

Posted by neptune.ml  over 1 year ago |  Quote

EDITED

Hi there,

We implemented the per pixel loss (still expeimenting)

We added deeper encoder architectures Resnet101, Resnet152 (still experimenting)

We added TTA up-down and left-right flips + rotation [0,90,180,270]

The best results we’ve gotten so far are trained without dice loss which results in not very smooth (though accurate) predictions. TTA smoothens then to a certain extend but we need some good models with dice to entirely get rid of that problem. Another option is better morphological post processing which we are also working on right now.

Everything that we do is as always available for you to use/change/comment in our open-solution repo https://github.com/minerva-ml/open-solution-mapping-challenge . Newest ideas are usually on the dev branch so make sure to check that out too.

Best, and good luck!