View in EDS

Dynamic focused prototypes distillation for few-shot object detection.

Saved in:
Bibliographic Details
Title: Dynamic focused prototypes distillation for few-shot object detection.
Authors: Xie, Binhong1 (AUTHOR) binhongxie@tyust.edu.cn, Wang, Rui1 (AUTHOR) ruiqz@stu.tyust.edu.cn, Pan, Lihu1 (AUTHOR) panlh@tyust.edu.cn
Source: Machine Vision & Applications. Nov2025, Vol. 36 Issue 6, p1-14. 14p.
Abstract: Existing meta-learning methods often struggle to generate accurate and information-dense prototypes due to the limited availability of novel class data, leading to decreased detection performance. Moreover, the Region Proposal Network (RPN) frequently fails to produce high-quality region proposals when faced with unstable feature representations and substantial intra-class feature variations. To address these challenges, we propose a novel few-shot object detection framework based on Dynamic Focused Prototypes Aggregation (DFPA). DFPA introduces a Bi-Level Routing Attention mechanism to extract prototypes from the support set, ensuring that the generated class-level prototypes are rich in detail. These prototypes are then aggregated into the feature map based on similarity, thereby enhancing feature representation before region proposal generation. In addition, we propose a Multi-Scale Convolutional Gated (MSCG) module, designed to improve the model’s sensitivity to targets of varying scales by capturing multi-scale features. MSCG also employs a convolutional gated mechanism to mitigate the negative effects of intra-class feature variations on the RPN. Experimental results demonstrated that on three novel class splits of the PASCAL VOC, the nAP50 of our method was improved by up to 2.3 percentage points compared with the baseline (Variational Feature Aggregation). Under the 30-shot setting of the MS COCO, nAP was increased by 0.8 percentage point compared to the meta-learning method SMPCCNet (Support-query Mutual Promotion and Classification Correction Network), achieving significant accuracy improvement in few-shot object detection. [ABSTRACT FROM AUTHOR]
Copyright of Machine Vision & Applications is the property of Springer Nature and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database: Engineering Source
Description
Abstract:Existing meta-learning methods often struggle to generate accurate and information-dense prototypes due to the limited availability of novel class data, leading to decreased detection performance. Moreover, the Region Proposal Network (RPN) frequently fails to produce high-quality region proposals when faced with unstable feature representations and substantial intra-class feature variations. To address these challenges, we propose a novel few-shot object detection framework based on Dynamic Focused Prototypes Aggregation (DFPA). DFPA introduces a Bi-Level Routing Attention mechanism to extract prototypes from the support set, ensuring that the generated class-level prototypes are rich in detail. These prototypes are then aggregated into the feature map based on similarity, thereby enhancing feature representation before region proposal generation. In addition, we propose a Multi-Scale Convolutional Gated (MSCG) module, designed to improve the model’s sensitivity to targets of varying scales by capturing multi-scale features. MSCG also employs a convolutional gated mechanism to mitigate the negative effects of intra-class feature variations on the RPN. Experimental results demonstrated that on three novel class splits of the PASCAL VOC, the nAP50 of our method was improved by up to 2.3 percentage points compared with the baseline (Variational Feature Aggregation). Under the 30-shot setting of the MS COCO, nAP was increased by 0.8 percentage point compared to the meta-learning method SMPCCNet (Support-query Mutual Promotion and Classification Correction Network), achieving significant accuracy improvement in few-shot object detection. [ABSTRACT FROM AUTHOR]
ISSN:09328092
DOI:10.1007/s00138-025-01761-1