Towards OOD Object Detection With Unknown-Concept Guided Feature Diffusion.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Humans possess a strong ability to learn knowledge about known objects.
• Unknown objects often deviate from established knowledge, posing challenges for detection systems.
• Unsupervised out-of-distribution object detection (OOD-OD) aims to identify novel objects without labeled OOD data.

Purpose of the Study:

• To develop a method for reasoning unknown concepts from existing knowledge.
• To enable unsupervised OOD-OD by synthesizing virtual OOD features.
• To address the challenge of detecting unseen objects without auxiliary OOD data during training.

Main Methods:

• Proposing Unknown-Concept Guided Feature Diffusion (UCFD).
• Utilizing an object-related knowledge extractor with learnable codewords to capture visual knowledge and enhance in-distribution (ID) object features.
• Constructing an unknown-concept pool by mixing codewords and employing an unknown-concept guided diffusor to generate virtual OOD features.

Main Results:

• Demonstrated significant performance gains on three different OOD-OD tasks.
• Showcased the ability to synthesize effective virtual OOD features through extensive visualizations.
• Validated the superiority of the proposed UCFD method in unsupervised OOD-OD scenarios.

Conclusions:

• UCFD effectively leverages learned knowledge to reason and synthesize unknown concepts for OOD-OD.
• The method successfully generates virtual OOD features, mitigating the need for OOD training data.
• UCFD offers a promising approach for improving the detection of unseen objects in real-world applications.