A new arXiv preprint, Preserving the Manifold: Frozen Expert Alignment for Label-Efficient SAR Scene Understanding, addresses the problem of aligning synthetic aperture radar (SAR) imagery with natural language. According to the abstract, this cross-modal alignment is held back by two distinct failures.
The first failure is modality mismatch: text encoders pre-trained on optical imagery, such as CLIP, cannot be directly transferred to SAR data. The abstract is cut off before describing the second failure or the specifics of the proposed method, but the title points to a "Frozen Expert Alignment" strategy intended to preserve the data manifold while enabling label-efficient scene understanding.
Because the source is truncated, this summary can only report the first identified failure and the paper's stated goal. No details about the second failure or the method's performance are available from the source provided.