[Q] Possible to merge overlapping tiles of a multi-class discrete model output using max strategy?

[cwerner]

Hi 👋

I just found your library as I was browsing for a tiler/ merger solution to replace my crappy custom code that cannot handle overlap.

I'm looking for a library that lets me split a source images (3-channel, 2048x2048px) into overlapping 3-channel tiles with size 256x256px (overlap width=32px in x and y direction).

I'm feeding my segmentation model batches of those tiles. The results are 1-channel images of size 256x256 (values are multi-class integer results: currently classes 0,1,2).

So the flow is:

1. Source: 2048x2048 RGB image
2. Convert to batches of 256x256 RGB tiles (that overlap by 32px): 9x9 tiles in total (with Tiler()?), non-overlapping it is 8x8 tiles
3. Run my model with those tiles
4. The model returns 81x 256x256 data tiles
5. Use merger to reconstruct a 2048x2048 monochrom image (with discrete class ids: 0, 1, 2) - in the overlap regions I would need to use the max value (no average or blending)

If I understand the library correctly, overlapping tiles are merged and the values are averaged? In my case, I'd have to take the max value in an overlap area. Is this supported? Could I adapt the code to take the max instead of the mean? And if so: are there any examples how to do this?

Cheers and thanks for any hints

[cwerner]

I just worked around it be simply adding the raw model logits and doing the argmax after merging everything with tiler. So this can be closed I guess...

[the-lay]

Hi, apologies for taking so long to reply!
Glad you have made it work!

[cwerner]

No problem 😉 nice little library by the way - just what I needed 👍