QSNA: Base-Conditioned Low-Rank Residual Refinement for Retention-Constrained Open-Vocabulary Detection on ROADWork

Adapting an open-vocabulary detector to work-zone imagery can improve target recognition but may reduce performance on source-domain and large-vocabulary benchmarks, represented here by COCO and LVIS. Replay and model merging can produce retention-aware checkpoints, but they do not by themselves provide a dedicated local refinement step after checkpoint selection. We propose Query Spectral-Null Adapter (QSNA), a complementary local update applied after checkpoint selection. QSNA freezes the selected checkpoint and learns an 8064-parameter low-rank residual at a single decoder FFN output projection, constrained by an estimated source-gradient complement. The residual can be rescaled after training and folded directly into the checkpoint. On the primary replay-derived checkpoint, QSNA raised ROADWork AP from 28.18 to 28.30, a +0.12 AP change, while COCO and LVIS AP decreased by 0.09 and 0.16 relative to the base. Across three fixed-base adapter seeds, the primary configuration produced similar descriptive ROADWork changes of about +0.13 AP and all three seeds remained above the adopted retention floors; this is replication evidence, not a significance test. Across nine base-specific fits, ROADWork AP repeatedly rose over a finite residual-scale range and then turned over, with the feasible range governed by the starting checkpoint. Canonical random, unconstrained, and projection-removed controls show that update orientation changes the acquisition–retention response, but they do not establish spectral projection as necessary or universally superior. Together, these results position QSNA as a base-conditioned, deployment-oriented local refinement stage after retention-aware checkpoint construction.