Sequential Instance Refinement for Cross-Domain Object Detection in Images.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Cross-domain object detection adapts models to new datasets.
• Existing methods struggle with domain shift due to outlier and irrelevant instances.
• Direct feature alignment can lead to negative transfer.

Purpose of the Study:

• To develop a robust method for cross-domain object detection.
• To address the challenges of outlier target and low-relevance source instances.
• To improve the performance of object detection models across different domains.

Main Methods:

• Proposes a reinforcement learning-based method named sequential instance refinement.
• Employs two agents to progressively refine source and target instances.
• Utilizes sequential actions to remove outlier target and low-relevance source instances.

Main Results:

• Demonstrates superior performance over state-of-the-art baselines.
• Achieves significant improvements on several benchmark datasets.
• Effectively handles outlier target and low-relevance source instances.

Conclusions:

• The proposed sequential instance refinement method enhances cross-domain object detection.
• Reinforcement learning offers a promising direction for addressing domain shift challenges.
• The method effectively mitigates negative transfer for more accurate object detection.