Toward Adaptive Open-Set Object Detection via Category-Level Collaboration Knowledge Mining.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Existing object detection methods face challenges in generalizing across diverse data domains and adapting to new categories.
• Adaptive Open-Set Object Detection (AOOD) aims to address this by using supervised learning on known categories and unsupervised adaptation for both known and novel categories.
• Current AOOD methods suffer from poor cross-domain feature representation, ambiguity in novel classes, and bias towards the source domain.

Purpose of the Study:

• To propose a category-level collaboration knowledge mining strategy to overcome limitations in existing AOOD approaches.
• To enhance cross-domain feature representation and reduce inter-category ambiguity for novel classes.
• To mitigate source domain bias in adaptive open-set object detection.

Main Methods:

• Developed a clustering-based memory bank (CMB) to integrate prototype, auxiliary, and intra-class disparity features for rich category-level knowledge.
• Introduced a base-to-novel selection metric (BNSM) to identify and leverage source domain features for novel categories.
• Implemented an adaptive feature assignment (AFA) strategy for transferring knowledge to the target domain and mitigating source domain bias through asynchronous memory bank updates.

Main Results:

• The proposed category-level collaboration knowledge mining strategy significantly improves performance in adaptive open-set object detection.
• Experiments show consistent outperformance over state-of-the-art AOOD methods across diverse datasets.
• Achieved performance gains ranging from 1.1 to 5.5 mean Average Precision (mAP).

Conclusions:

• The proposed method effectively addresses the limitations of existing AOOD approaches by enhancing cross-domain generalization and adaptation to novel categories.
• The category-level collaboration knowledge mining strategy provides a robust framework for exploiting inter- and intra-class feature relationships.
• The approach demonstrates superior performance and offers a promising direction for future research in adaptive open-set object detection.