COSOS-1k: A Benchmark Dataset and Occlusion-Aware Uncertainty Learning for Multi-View Video Object Detection

Area of Science:

• Computer Vision
• Occupational Safety
• Machine Learning

Background:

• Confined spaces present significant risks to workers due to their enclosed nature.
• Evaluating confined space operational safety (COSOS) is critical but challenged by crowded environments and small equipment.
• Existing research lacks dedicated datasets for COSOS tasks.

Purpose of the Study:

• To introduce COSOS-1k, the first dataset tailored for real-world COSOS scenarios.
• To address occlusion challenges in confined space safety evaluations.
• To propose novel methods for improving object detection and attribute recognition under occlusion.

Main Methods:

• Construction of the COSOS-1k dataset with multi-view videos, 10 safety equipment types, 6 worker attributes, and detailed annotations.
• Development of Occlusion-aware Uncertainty Estimation (OUE) for part-level occlusion prediction.
• Introduction of Cross-Frame Cluster (CFC) and Cross-View Cluster (CVC) attention mechanisms to handle temporal and view-based occlusions.

Main Results:

• The COSOS-1k dataset enables diverse and expressive research in confined space safety.
• Proposed OUE, CFC, and CVC methods effectively mitigate occlusion issues.
• Experiments validate the effectiveness of the new methods and highlight dataset importance.

Conclusions:

• The COSOS-1k dataset and proposed methods significantly advance research in confined space operational safety.
• Addressing occlusion is crucial for accurate safety equipment recognition in complex environments.
• Dataset diversity and expressivity are key for developing robust computer vision models for occupational safety.