Object detectors in street scenes tend to silently miss out-of-distribution (OOD) objects, producing no detection and leaving downstream systems unaware that anything was missed. Existing OOD methods mostly target image classification or rely on auxiliary branches, and they rarely handle in-distribution and OOD objects in a unified way. The authors propose SynOE-OD, a synthetic outlier-exposure approach that uses text-guided inpainting to create synthetic outliers offline and an open-vocabulary object detector to verify them.

These synthetic examples are then used to fine-tune an object detector with an additional OOD class. At inference, the detector needs no auxiliary branch, so the extra OOD awareness comes without added architectural complexity. On the OoDIS benchmark, SynOE-OD outperforms the strongest open-vocabulary baseline, reaching AP_50:95 of 57.6% on RoadObstacle, 29.4% on RoadAnomaly, and 39.2% on Fishyscapes, compared with 48.5%, 22.7%, and 22.7% for the baseline.

The reported gains hold for both fine-tuned closed-set and open-vocabulary detectors, while in-distribution performance in street scenes stays competitive. The paper is a single source; no independent comparison or replication is described in the source material.