Self-Supervised Multi-Category Counting Networks for Automatic Check-Out.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Automatic Check-Out (ACO) systems face challenges with large-scale, fine-grained product categories and continuous product updates.
• Existing ACO methods often rely on labor-intensive bounding box annotations, limiting scalability and adaptability.
• Realistic check-out scenarios require systems that can handle evolving product inventories.

Purpose of the Study:

• To propose a novel Self-Supervised Multi-Category Counting (S2MC2) network for Automatic Check-Out (ACO).
• To reduce the cost of labeling by utilizing point-level product supervision.
• To enable ACO predictions in a class-incremental setting, accommodating new products over time.

Main Methods:

• Developed the S2MC2 network with a class-agnostic counting backbone.
• Incorporated an attention module for fine-grained pattern recognition.
• Integrated a domain adaptation module to bridge the gap between training and testing data domains.
• Employed a self-supervised approach for backbone parameter initialization.

Main Results:

• The S2MC2 network demonstrated superior accuracy on the RPC dataset for ACO tasks.
• Achieved strong performance in both traditional and class-incremental settings.
• Outperformed competing baseline methods in comprehensive experiments.

Conclusions:

• The proposed S2MC2 network offers an effective solution for Automatic Check-Out challenges.
• Leveraging self-supervised learning and point-level supervision significantly reduces labeling effort.
• The S2MC2 network provides a robust and adaptable approach for real-world, dynamic check-out environments.