Radar-camera depth estimation aims to convert sparse but metric radar measurements into dense depth maps. According to a new arXiv paper, existing methods typically combine radar and image features outside the sequence operator of the backbone, often through concatenation.

The paper proposes an alternative: radar-modulated state space models. Instead of fusing features after the fact, the radar signal is used to modulate the state space dynamics inside the sequence operator. The title—'Selection, Not Fusion'—captures this shift from feature concatenation to selective modulation.

The abstract does not include experimental results, so the effectiveness of the approach remains to be seen. The work is available on arXiv under identifier 2605.11840.