Conditional Variational Memory-Augmented Aggregation for Few-Shot Object Detection

Abstract

Few-shot object detection (FSOD) aims to build a universal detector that performs reliably with limited annotations and has gained increasing attention for its potential to reduce data reliance. However, existing approaches generally rely on a fixed support feature to assist in detecting different query instances within the same image, limiting the model’s adaptability to query variations. To address this issue, we propose a novel meta-learning framework, Conditional Variational Memory-Augmented Aggregation Network (CVMA). The framework introduces the Conditional Variational Guidance Module (CVGM), which incorporates a conditional variational encoder to dynamically generate support features. This module effectively mitigates the inherent bias toward base classes by enhancing the compatibility of the support features with query instances. Secondly, we propose a Memory-Augmented Aggregation Module (MAM), which leverages a memory mechanism to perform dynamic aggregation between optimized support and query features, enabling more effective mining of semantic correlations and task adaptability. Extensive experiments on PASCAL VOC and COCO demonstrate that CVMA consistently outperforms state-of-the-art methods across various few-shot settings, showing superior detection performance.

Keywords

Few-Shot Object Detection / Meta-Learning / Memory Network

Rights & permissions

Higher Education Press 2026

Just Accepted

This article has successfully passed peer review and final editorial review, and will soon enter typesetting, proofreading and other publishing processes. The currently displayed version is the accepted final manuscript. The officially published version will be updated with format, DOI and citation information upon launch. We recommend that you pay attention to subsequent journal notifications and preferentially cite the officially published version. Thank you for your support and cooperation.