PixOOD: Pixel-Level Out-of-Distribution Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Out-of-distribution (OOD) detection is crucial for reliable AI systems.
• Traditional OOD methods often require anomalous data samples for training, leading to biases.
• Pixel-level OOD detection faces challenges due to complex intraclass data variability.

Purpose of the Study:

• To introduce PixOOD, a novel pixel-level OOD detection algorithm.
• To develop an OOD detection method that does not require anomalous data for training.
• To create a versatile algorithm applicable to various domains without specific application design.

Main Methods:

• PixOOD utilizes an in-distribution data model and a decision strategy estimator.
• An online data condensation algorithm models pixel-level intraclass variability robustly.
• Per-class and unified calibration models are proposed for decision strategy estimation.

Main Results:

• PixOOD achieved state-of-the-art performance on four out of seven diverse datasets.
• The algorithm demonstrated competitive results on the remaining datasets.
• The proposed online data condensation is more robust than K-means and trainable via gradient descent.

Conclusions:

• PixOOD offers a robust and unbiased approach to pixel-level OOD detection.
• The method's flexibility makes it suitable for various applications.
• The developed algorithm pushes the state-of-the-art in OOD detection research.