WHY MAE

Solid work!

I am curious about the selection of image encoder. Why did you adopt MAE as your image encoder? Why not DINO?

In my opinion, semantic segmentation heavily rely on semantic information, but MAE is trained as an autoencoder,
which intuitively lacked of semantic information.

In contrast, DINO feature is widely used, which can be used to predict semantic map, depth map, normal map.....

So, why not DINO? Why MAE?

Thanks

The issue names no files, tests, or entry points to inspect. Start by reviewing the repository's image-encoder and training documentation, if available, and identify where the MAE choice is explained. Done would require a clear, maintainer-approved rationale comparing MAE with DINO, but the issue does not define a documentation location or an expected format.

Written by the indexing model from the issue text.