Unsupervised Out-of-Distribution Object Detection via PCA-Driven Dynamic Prototype Enhancement.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object detectors struggle with out-of-distribution (OOD) data, leading to overconfident predictions on unseen objects.
• Supervision for unknown data is lacking, hindering the development of robust OOD object detection (OOD-OD).

Purpose of the Study:

• To develop a method for synthesizing OOD data to improve object detector performance on unseen objects.
• To enhance the localization and discrimination capabilities of object detectors for OOD data.

Main Methods:

• Proposed a PCA-Driven dynamic prototype enhancement method to extract simulative OOD data.
• Utilized principal components to create an OOD map and dynamic prototypes for improved discrimination.
• Defined a contrastive loss to increase the semantic gap between OOD and in-distribution data.

Main Results:

• Demonstrated significant performance gains on OOD-OD and incremental object detection tasks.
• The method effectively alleviates the impact of lacking unknown data during training.
• Showcased improved ability to discriminate OOD objects.

Conclusions:

• PCA-Driven dynamic prototype enhancement offers a superior approach for OOD object detection.
• The proposed method enhances robustness by effectively handling unseen data.
• This work contributes to more reliable object detection in real-world applications.