TernausNetV2

TernausNetV2 copied to clipboard

Hello, I have a question: how to predict areas of an image (binary mask) where different objects touch or very close to each other?

Hello @ternaus ,

Thank you very much for sharing your solution, very clean work.

I have a similar question so I'm posting here. I hope you won't be too busy with Lyft's Kaggle competition!

I see that the network outputs two values per pixel, one for semantic segmentation (the binary mask) and one for the instance borders. But does this mean that you need to have explicit labels for both of them? Or did you use some preprocessing technique in order to create borders from binary masks? (by looking at the challenge website it seems that you only had binary masks but I may have misread something)

Bests!

For the data, we have instance-level masks.

During the processing step I convert them into binary masks and regions where different buildings touch each other.

Later I use both of them as a target for training.

hi @ternaus, can you share the code to calculate the touching mask from binary building mask? Thanks in advance!

hi @ternaus, can you share the code to calculate the touching mask from binary building mask? Thanks in advance!

Hi sir, I wonder that have you gotten the code. If not, I can share with you.