基于区域的上下文信息矩阵推理目标检测模型
Contextual Matrix for Reasoning-based Object Detection
邵明文 1范冰冰 1彭子路 1李云昊1
作者信息
- 1. 中国石油大学(华东)计算机科学与技术学院,山东 青岛 266580
- 折叠
摘要
目标的语义信息或目标之间的关系有助于目标识别.然而,目前最先进的目标检测器依赖于单独识别目标实例,缺乏充分的上下文信息.为了充分利用上下文信息,本文提出了一种特定的上下文矩阵,将上下文矩阵结合到目标检测中来增强检测模型的上下文信息推理能力.具体来说,该上下文矩阵有 3 种形式,分别来自不同数据集的标注.通过目标之间的关系、属性和共现概率来增强上下文信息,并将每个区域的上下文矩阵增强后的新特征与原始特征相连接,提高分类和定位的性能.此外,基于上下文矩阵推理的模型是轻量级的和灵活的,足以增强不同的目标检测基线并有助于少样本检测任务.大量实验表明,所提出的基于上下文矩阵推理的检测器可以在不同的基准上持续改进各种检测器.
Abstract
It has been well recognized that contextual information or relation between objects helps object recognition.However,the most of current state-of-the-art object detectors rely on recognizing object instances individually and lack contextual infor-mation.To make full use of the contextual information,we proposed a specific contextual matrix to enhance the context infor-mation reasoning ability of the detection model by incorporating the contextual matrix in object detection.Specifically,the con-text matrix has three forms which obtained from the annotations of different data sets,to enhance the context information through the relationship,attributes and the co-occurrence probability of the objects.Further,the contextual matrix reasoning-based model was lightweight and flexible enough to enhance different object detection baseline models and help few-shot task.Extensive experiments illustrate that the proposed contextual matrix reasoning-based detector can consistently improve various detectors on different benchmarks.
关键词
深度学习/目标检测/上下文矩阵/视觉推理Key words
deep learning/object detection/contextual matrix/visual reasoning引用本文复制引用
基金项目
国家自然科学基金项目(62356285)
山东省自然科学基金项目(ZR2022MF260)
出版年
2024
NETL
NSTL
万方数据