Proteus
Learn across modalities. Sense with radar.
Transferring visual knowledge to mmWave representations, with more informative radar imaging.
A second modality can shape learning without becoming a permanent deployment requirement.
The question
Can a radar recover useful wetness information without keeping a camera in the inference loop?
The idea
Reduce SAR speckle, incorporate phase-angle information, and train an mmWave student using an RGB teacher.
What the evidence shows
Proteus prototypes this approach on commercial radar and evaluates wetness classification across environmental scenarios. The teacher supports training; the student uses mmWave at inference. This is evidence for cross-modality transfer in this task, rather than a general claim of physical understanding.
Source: ACM SenSys 2025What it opens up
The work highlights the importance of what the radar representation preserves. It motivates my broader question: which physical distinctions can an incomplete observation actually support?
