Counting with ease: Class-agnostic counting via one-shot detection across diverse domains

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Class-agnostic counting is crucial for industrial and agricultural applications but current density map methods struggle with background interference and lack precise localization.
• Existing models face challenges due to suboptimal designs and a scarcity of datasets with bounding box annotations, hindering advancements.
• Text-guided methods, while explored, are impractical for edge deployment, necessitating new approaches.

Purpose of the Study:

• To propose a novel Class-Agnostic Counting and Localization (CACAL) framework for accurate and efficient object counting and localization.
• To address limitations of existing methods, including background interference and imprecise object locations.
• To introduce a comprehensive dataset for advancing class-agnostic counting research.

Main Methods:

• Developed a novel Class-Agnostic Counting and Localization (CACAL) framework utilizing a single query image.
• Introduced a Sampling-Aware Feature Enhancement module to improve feature discriminability in shared-encoder settings.
• Designed a Split-and-Assemble Feature Matching strategy for enhanced performance in cluttered and occluded scenarios.

Main Results:

• The proposed CACAL framework demonstrates consistent outperformance over existing methods across multiple benchmarks.
• CACAL achieves accurate object counting and localization, streamlining processes for real-world applications.
• The framework exhibits strong generalization capabilities across diverse industrial, agricultural, and daily-life domains.

Conclusions:

• The CACAL framework offers a significant advancement in class-agnostic counting and localization.
• The newly introduced LOCO dataset provides a large-scale benchmark with diverse annotations for future research.
• CACAL's performance and generalization pave the way for more robust and practical object counting solutions.