The scoring function mismatch: your OOD score may not fit your model
A second hidden cost is that your OOD scoring method must match how the detector outputs probabilities. Many detectors use a sigmoid function for multi-label classification, but classic OOD scores assume a softmax output. The paper highlights this incompatibility: using a softmax-based score on a sigmoid-based detector gives poor separation between in-distribution and OOD instances [2]. They solved it by designing a new sigmoid-based scoring function that uses both positive and negative labels, producing strong responses for known objects and suppressed responses for unknown ones [2]. The lesson: you can't just bolt on an off-the-shelf OOD score; you need to adapt it to your model's architecture, which is extra work and a potential source of failure if overlooked.
Training complexity: more moving parts, more tuning, and sometimes worse results
Structured knowledge methods often add complex training components that can backfire if not tuned carefully. One approach uses learnable codewords to extract object knowledge and a diffusion process to synthesize virtual OOD features [1]. This adds significant complexity—you have to train the codewords, build an unknown-concept pool, and run a reverse diffusion process—all while avoiding degradation of in-distribution performance [1]. The authors report significant gains, but the complexity is a hidden cost in terms of compute and hyperparameter tuning. Worse, a separate study found that incorporating human-defined similarity rankings into contrastive learning actually hurt OOD detection performance [3]. That's a stark warning: not all structured prior knowledge is beneficial; some can actively mislead the model. The takeaway is that you need to validate whether your chosen prior actually helps, and be prepared for extra training overhead.
About These Sources
This answer is built on 3 studies (1 peer-reviewed, 2 preprints) — published from 2022 to 2026, 2 from 2024 or later, 1 in Q1 journals — selected as the most relevant from 3 studies that passed quality screening, drawn from 45 papers retrieved from a database of over 500 million.
Sources used in this answer
Towards OOD Object Detection With Unknown-Concept Guided Feature Diffusion
Proposes UCFD, which uses learnable codewords to extract object knowledge and a diffusion process to synthesize virtual OOD features, achieving significant performance gains across three tasks while noting the need to avoid degrading in-distribution detection.
NegAS: Negative Label Guided Attention and Scoring for Out-of-Distribution Object Detection with Vision-Language Models
Introduces NegAS for VLM-based detectors, using LLM-generated negative labels to guide attention and a sigmoid-based OOD scoring function, reducing FPR95 by 11.4% on COCO and 25.5% on OpenImages compared to baseline.
Contrastive Learning for OOD in Object detection
Shows that incorporating human-defined similarity rankings into contrastive learning can improve representation learning for classification and detection but actually hurts OOD detection performance, highlighting the risk of using biased priors.