MOODv2: Masked Image Modeling for Out-of-Distribution Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Effective out-of-distribution (OOD) detection requires robust in-distribution (ID) representations distinct from OOD samples.
• Prior methods often used recognition-based techniques, leading to shortcut learning and incomplete representations.

Purpose of the Study:

• To analyze the impact of different pretraining tasks and OOD score functions on detection performance.
• To develop an improved OOD detection framework leveraging effective pretraining strategies.

Main Methods:

• Conducted a comprehensive analysis of various pretraining tasks and OOD score functions.
• Employed reconstruction-based pretext tasks, specifically masked image modeling, for feature representation learning.
• Introduced the MOODv2 framework for OOD detection.

Main Results:

• Feature representations pre-trained via reconstruction significantly enhanced OOD detection performance.
• Reconstruction-based pretraining narrowed the performance gap between different OOD score functions.
• The MOODv2 framework achieved high AUROC scores: 95.68% on ImageNet and 99.98% on CIFAR-10.

Conclusions:

• Reconstruction-based pretext tasks are highly adaptable and effective for OOD detection.
• Simple OOD score functions can achieve competitive results when combined with strong reconstruction-based representations.
• The MOODv2 framework demonstrates the potential of masked image modeling for robust OOD detection.