Learning with Weak Annotations for Robust Maritime Obstacle Detection.

Area of Science:

• Computer Vision
• Autonomous Systems
• Machine Learning

Background:

• Maritime obstacle detection is crucial for autonomous navigation safety.
• Current methods rely on deep segmentation networks trained on large, expensively labeled datasets.
• Per-pixel ground truth labeling is a significant bottleneck.

Purpose of the Study:

• To introduce a novel scaffolding learning regime (SLR) for maritime obstacle detection.
• To reduce the labor-intensive and costly ground truth labeling effort.
• To train segmentation-based networks using weak annotations.

Main Methods:

• Leveraging weak annotations: water edges, horizon location, and obstacle bounding boxes.
• Training an initial model with weak annotations.
• Alternating between re-estimating segmentation pseudo-labels and network parameter refinement.
• Reducing annotation effort by a factor of twenty.

Main Results:

• Maritime obstacle segmentation networks trained with SLR and weak annotations outperform those trained with dense labels.
• SLR achieves comparable or superior accuracy to traditional methods.
• Enhanced domain generalization capabilities were observed.

Conclusions:

• Scaffolding learning regime (SLR) offers a highly effective alternative to dense labeling for maritime obstacle detection.
• SLR significantly reduces annotation costs while improving model performance and generalization.
• The approach facilitates domain adaptation with minimal manual annotation.