BROWSE
Detail
Conference Paper
SAM-OOD: Foundation-Model-Guided Unknown Mining for Object-Level Anomaly Detection
Cited - time in 
Share 
Share
- Authors
- Seher Kanwal, Seung-Ik Lee
- Issue Date
- 2026-06
- Citation
- Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) 2026, pp.7683-7692
- Publisher
- Computer Vision Foundation
- Language
- English
- Type
- Conference Paper
- Abstract
- Object-level anomaly detection under distribution shift requires an object detector to localize and recognize in-distribution categories while reliably flagging unknown objects in cluttered scenes. A key bottleneck is the scarcity of diverse unknown-labeled instances for training and evaluation. We propose SAM-OOD, a lightweight framework that leverages a class-agnostic vision foundation model as a training-time anomaly proposal generator to obtain scalable unknown supervision without introducing auxiliary OOD images or synthetic outliers. Specifically, we use Segment Anything Model (SAM) proposals to mine candidate regions from the original in-distribution training images and assign them to an explicit unknown class, enabling supervised K+1 detector training with no architectural modifications. At inference time, SAM is not used; we apply standard logit-based scoring (MSP or energy) to distinguish ID and OOD detections. We further identify an evaluation pitfall for K+1 detectors under incomplete annotations and propose probability-based filtering for reliable FPR95 threshold estimation. Experiments on Pascal VOC/COCO-style settings show that SAM-OOD achieves state-of-the-art object-level OOD detection, reaching 3.69% FPR95 and 99.04% AUROC, while preserving competitive in-distribution detection performance and incurring no additional inference-time overhead.
- KSP Keywords
- Lightweight framework, Model-guided, Object-level, Probability-based, Threshold Estimation, Training and evaluation, Unknown class, anomaly detection, art object, based filtering, detection performance