FAST: Foreground-aware active self-training for domain adaptive object detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Domain adaptive object detection (DAOD) trains models on unlabeled target data using labeled source data.
• Mean-teacher self-training is effective for DAOD but hindered by noisy pseudo-labels.
• Active domain adaptation (ADA) selectively annotates informative data to reduce costs, but its use in DAOD is limited.

Purpose of the Study:

• To develop an effective active learning framework for domain adaptive object detection.
• To address the challenge of noisy pseudo-labels in mean-teacher self-training for DAOD.
• To minimize annotation costs in DAOD by selecting the most valuable target samples.

Main Methods:

• Proposed Foreground-aware Active Self-Training (FAST) framework for active DAOD.
• Introduced foreground diversity clustering sampling (FDCS) to maximize foreground object diversity.
• Implemented teacher-student discrepancy uncertainty sampling (TDUN) to identify uncertain predictions.
• Utilized a decoupled active learning paradigm with a dedicated sampling model.

Main Results:

• FAST significantly enhances object detection performance on target domains.
• The proposed sampling strategies effectively identify informative target samples.
• Experimental results demonstrate superior performance across multiple DAOD datasets.
• The method successfully bridges the domain gap in challenging scenarios.

Conclusions:

• FAST establishes an effective framework for active domain adaptive object detection.
• The combination of FDCS and TDUN sampling strategies optimizes annotation efforts.
• The proposed approach offers a promising solution for improving DAOD with minimal annotation cost.