Domain adaptive object detection with model-agnostic knowledge transferring.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Convolutional Neural Network (CNN)-based object detectors have advanced significantly due to deep learning.
• Performance degrades in new scenarios due to distribution variance between training and testing domains.
• Labeling new data is costly and time-consuming, necessitating effective domain adaptation methods.

Purpose of the Study:

• To address performance degradation in object detection caused by domain shift.
• To propose a novel domain adaptation method that avoids negative feature transfer.
• To enhance the recognition ability of object detectors in unlabeled target domains.

Main Methods:

• Introduced the Knowledge Transfer Network (KTNet) with two core modules: object intrinsic knowledge mining and category relational knowledge constraint.
• Utilized a shared binary classifier for adaptive alignment of foreground and background features across domains.
• Constructed relational knowledge graphs to explicitly constrain category correlations in source, target, and cross-domain settings.

Main Results:

• KTNet effectively guides detectors to learn object-related and domain-invariant representations.
• The proposed method achieved strong performance in four common cross-domain scenarios.
• Ablation experiments confirmed the method's scalability with complex backbone networks and detection architectures.

Conclusions:

• KTNet offers a robust solution for domain adaptation in object detection, overcoming limitations of adversarial training.
• The approach successfully aligns features and constrains category relationships, improving cross-domain recognition.
• The proposed modules are effective and adaptable to various deep learning-based object detection frameworks.