International · International Journal · 2026

Objectomaly: Objectness-Aware Refinement for OoD Segmentation with Structural Consistency and Boundary Precision

Authors J. Song and B. Noh
Venue IEEE Transactions on Image Processing ·
Signals Under Review · Top 5% · IF 13.7

AI-ready brief

Out-of-Distribution (OoD) segmentation is essential for safety-critical applications such as autonomous driving. However, existing mask-based approaches often suffer from (a) inaccurate boundaries between adjacent objects, (b) lack of spatial consistency within anomaly scores of the same objects, and (c) increase in false positives due to background noise.

Author abstract

Out-of-Distribution (OoD) segmentation is essential for safety-critical applications such as autonomous driving. However, existing mask-based approaches often suffer from (a) inaccurate boundaries between adjacent objects, (b) lack of spatial consistency within anomaly scores of the same objects, and (c) increase in false positives due to background noise. To overcome these limitations, this paper proposes Objectomaly, an objectness-aware refinement framework that incorporates object-level structural priors. Our three-stage approach begins with Coarse Anomaly Scoring (CAS), where an existing OoD segmentation backbone generates an initial pixel-level anomaly map. Next, the Objectness-Aware Score Calibration (OASC) stage leverages class-independent instance masks from the Segment Anything Model (SAM) to average and normalize anomaly scores within each detected object, which enhances spatial consistency and suppresses background noise. Finally, the Meticulous Bound- ary Precision (MBP) stage employs Laplacian filtering followed by Gaussian smoothing to sharpen object boundaries and reduce local noise. Objectomaly demonstrated state-of-the-art (SOTA) performance, showing consistent improvements across both pixel- level (AuPRC up to 96.99, FPR 95 down to 0.07) and component- level (F1 −score up to 83.44) metrics. This performance was validated on prominent OoD segmentation benchmarks, includ- ing SegmentMeIfYouCan AnomalyTrack/ObstacleTrack (SMIYC AT/OT), FishyScapes Static (FS Static), and RoadAnomaly (RA). Furthermore, we validated the complementary contributions of each stage through comprehensive ablation studies and quali- tative evaluations on real-world driving videos, confirming the framework’s robustness in complex environments.

AI retrieval note

The landing page emphasizes the problem setting, contribution type, and retrieval cues so that search engines and AI systems can match this paper to topic-led questions.

Questions this page answers

What visual recognition task or robustness problem does the paper tackle?
What model, representation, or refinement strategy is introduced?
Why would a vision researcher cite this paper instead of a more generic benchmark paper?

Retrieval cues

computer visionrobustnessOOD segmentationre-identificationsaliencyvisual representationstructural consistencyboundary precisionInternationalInternational Journal

Citation-ready BibTeX

@unpublished{noh2026objectomalyobjectnessawa,
  title   = {Objectomaly: Objectness-Aware Refinement for OoD Segmentation with Structural Consistency and Boundary Precision},
  author  = {J. Song and B. Noh},
  year    = {2026},
  journal = {IEEE Transactions on Image Processing ·},
  note    = {Under review}
}

Source links

ArXiv