SQLNet: Scale-Modulated Query and Localization Network for Few-Shot Class-Agnostic Counting.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Class-agnostic counting (CAC) aims to count objects of any class using exemplar images.
• Existing methods often use density map regression, limiting object localization and scale exploration.
• Current approaches inefficiently process exemplars individually, hindering information synthesis.

Purpose of the Study:

• To propose a novel localization-based CAC approach, SQLNet, addressing limitations of existing methods.
• To improve object counting accuracy by incorporating exemplar scale information.
• To enable precise object localization and size prediction for downstream applications.

Main Methods:

• SQLNet utilizes a novel query and localization strategy, fully exploring exemplar scales.
• The Hierarchical Exemplars Collaborative Enhancement (HECE) module extracts discriminative features using multi-scale exemplar cooperation.
• The Exemplars-Unified Query Correlation (EUQC) module unifies exemplar and query feature interaction.
• The Scale-aware Multi-head Localization (SAML) module predicts object confidence, location, and size.
• A scale-aware localization loss enhances supervision using flexible location associations and exemplar scales.

Main Results:

• SQLNet demonstrates superior performance compared to state-of-the-art methods on CAC benchmarks.
• The method achieves excellent accuracy in object counting.
• SQLNet also excels in precise object localization and bounding box generation.

Conclusions:

• SQLNet offers a significant advancement in class-agnostic counting by integrating localization and scale awareness.
• The proposed approach overcomes the limitations of density map regression methods.
• SQLNet provides a practical and effective solution for tasks requiring object counting, localization, and size estimation.