Conditional Prototype Learning for Few-Shot Object Detection


Zhenwei He (Chongqing University of Technology), Xinye Liao (Chongqing University of Technology), Xin Feng (Chongqing University of Technology)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)

Abstract

Few-Shot Object Detection (FSOD) aims to train models for detecting novel objects by leveraging abundant data from base classes and only a few examples from novel classes. In the meta-learning paradigm for FSOD, the class prototypes are learned to guide the feature learning. However, The particularity of limited samples may lead prototypes to fit specific instances — semantic bias, which compromises the generalization ability. To address this, we propose Conditional Prototype Learning (CPL), a novel framework that integrates robust conditional references with training data to mitigate bias and enhance prototype robustness. Specifically, we introduce a Masked Conditional Variational Autoencoder (MCVA) to generate more robust conditional prototypes by correcting semantic biases. Furthermore, since ideal prototypes should not retain sample-specific particularities, the conditional prototypes lack explicit spatial and scale information. We design the task-aware feature aggregation (TFA) module to enhance the feature presentation for classification and regression tasks. Extensive experiments on PASCAL VOC and MS COCO demonstrate that our approach achieves state-of-the-art performance, validating the effectiveness of CPL in improving generalization to novel classes. Our code is now available at https://github.com/yefeng23323/CPL

Citation

@inproceedings{He_2025_BMVC,
author    = {Zhenwei He and Xinye Liao and Xin Feng},
title     = {Conditional Prototype Learning for Few-Shot Object Detection},
booktitle = {36th British Machine Vision Conference 2025, {BMVC} 2025, Sheffield, UK, November 24-27, 2025},
publisher = {BMVA},
year      = {2025},
url       = {https://bmva-archive.org.uk/bmvc/2025/assets/papers/Paper_1051/paper.pdf}
}


Copyright © 2025 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection