DecAlign — Reading Notes

Why I read this paper

I am interested in a recurring question in multimodal learning: what should be shared across modalities, and what should remain modality-specific?

Problem

Heterogeneous sensors observe complementary properties of the same physical world, so forcing every feature into a fully shared space may discard useful modality-specific information.

What I want to remember

• Alignment is a design choice, not a universal objective.
• Shared representations should be evaluated together with modality-specific information retention.
• A useful fusion mechanism should answer where to align, what to preserve, and when to exchange information.

My thoughts

For RGB–LiDAR perception, image appearance and point-cloud geometry are correlated but not interchangeable. A useful representation should support cross-modal correspondence without erasing complementary structure.