Domain adaptive underwater object detection via complementary style-aware learning.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Underwater images present multi-style variations due to water quality, lighting, and imaging devices, causing a domain gap.
• This domain gap, coupled with limited annotated data, degrades object detection performance in cross-domain tasks.
• Existing Mean Teacher frameworks struggle with pseudo-label quality, limiting their effectiveness in cross-domain object detection.

Purpose of the Study:

• To propose a complementary style-aware Mean Teacher (CSAMT) model to enhance cross-domain object detection performance.
• To address the limitations of pseudo-label generation in traditional Mean Teacher frameworks for underwater imagery.
• To improve the robustness and accuracy of object detection in diverse underwater environments.

Main Methods:

• Constructing image pairs for style-aware learning and performing style-content disentanglement using WCT2.
• Leveraging Discrete Wavelet Transform (DWT) for joint feature modeling across frequency and spatial domains.
• Introducing a two-stage teacher-student region proposal alignment (TTRPA) strategy for improved attention and consistency loss.

Main Results:

• The proposed CSAMT model achieves state-of-the-art performance on three domain adaptation benchmarks.
• Ablation studies confirm the significant contribution of each component of the CSAMT model.
• The model demonstrates improved object detection accuracy in challenging cross-domain underwater scenarios.

Conclusions:

• The CSAMT model effectively bridges the domain gap in underwater object detection by improving pseudo-label quality.
• Style-aware learning and TTRPA are crucial for enhancing the performance of Mean Teacher frameworks in cross-domain settings.
• This research offers a promising approach for robust object detection in diverse and challenging underwater environments.