泛研网

公众版 查看说明
全球科研项目数据库2026版

面向服务机器人的无监督领域自适应目标检测方法研究

项目来源

国家自然科学基金(NSFC)

项目主持人

叶茂

项目受资助机构

电子科技大学

项目编号

61773093

立项年度

2017

立项时间

未公开

研究期限

未知 / 未知

项目级别

国家级

受资助金额

66.00万元

学科

信息科学-人工智能-机器感知与机器视觉

学科代码

F-F06-F0604

基金类别

面上项目

关键词

深度神经网络 ; 回复式神经网络 ; 目标检测 ; 生成对抗网络 ; Object Detection ; Deep Neural Network ; Recurrent Neural Network ; Generative Adversarial Nets

参与者

李凡；邢冠宇；徐培；任东晓；李旭东；唐宋；张锋；刘丹；淦艳

参与机构

电子科技大学；四川大学；浙江科技学院；上海理工大学；南京邮电大学；重庆大学

项目标书摘要：当服务机器人工作于新场景时，因源场景与新场景数据分布的不一致，目标检测效果通常会下降。迁移学习是一个很好的解决手段，但目前绝大多数领域自适应目标检测方法要求在迁移的时候保留源样本，或者少量新场境样本有标签。这些要求服务机器人通常无法满足。针对这种情况，在以往深度学习和机器视觉研究基础上，本课题提出通过网络调控和一致特征学习等方法实现目标检测器对新场景的无监督迁移。研究内容有：基于网络调控的目标检测神经网络迁移方法研究；基于无监督一致特征学习的目标检测神经网络迁移方法研究；结合上下文信息与融入多模态信息的无监督领域自适应目标检测方法研究。创新之处有：不保留源训练集、没有新场景目标标签，基于无监督学习的目标检测网络迁移模型；基于调控网络的迁移学习框架；基于网络调控和无监督学习的上下文和多模态信息的融合方法。取得的成果不仅能丰富目标检测方法和机器学习理论，也有很重要的社会与经济价值。

Application Abstract: When a service robot is working in a new scene,because of the inconsistency of the data distributions between the source and the new scenes,the detection performance of object detector always will drop rapidly.Transfer learning is a good solution;however,almost all of present domain adaption methods require that the training set is kept or some of target samples are labeled.These requirements for service robots are usually not met.From the previous research experiences in the fields of deep learning and machine vision,we propose to study the unsupervised transfer methods which are based on network control and consistent feature learning.The research contents include:1network control framework for transferring neural network based object detector,2unsupervised consistent feature learning model for transferring neural network based object detector,3unsupervised context feature learning methods and the corresponding object detector transferring methods,4unsupervised multi-modal information learning methods and the corresponding object detector transferring methods.The main innovation points are:1network control framework for transferring learning,2unsupervised consistent feature learning model,3the combing methods which absorb context and multi-modal information based on network control and unsupervised learning.The results obtained through these studies are expected not only to enrich the methods of object detection of robot and machine learning theory,but also to make important contributions to social and economic development.

项目受资助省

四川省

项目结题报告

面向服务机器人的无监督领域自适应目标检测方法研究结题报告(全文)

  • 排序方式：
  • 1
  • /

  • 1.Bidirectional generative transductive zero-shot learning

    • 关键词：
    • Zero-shot learning; Transductive; Bidirectional generation; CycleGAN

    Most zero-shot learning (ZSL) methods aim to learn a mapping from visual feature space to semantic feature space or from both visual and semantic feature spaces to a common joint space and align them. However, in these methods the visual and semantic information are not utilized sufficiently and the useless information is not excluded. Moreover, there exists a strong bias problem that the instances from unseen classes always tend to be predicted as some seen classes in most ZSL methods. In this paper, combining the advantages of generative adversarial networks (GANs), a method based on bidirectional projections between the visual and semantic feature spaces is proposed. GANs are used to perform bidirectional generations and alignments between the visual and semantic features. In addition, cycle mapping structure ensures that the important information are kept in the alignments. Furthermore, in order to better solve the bias problem, pseudo-labels are generated for unseen instances and the model is adjusted along with them iteratively. We conduct extensive experiments at traditional ZSL and generalized ZSL settings, respectively. Experiment results confirm that our method achieves the state-of-the-art performances on the popular datasets AWA2, aPY and SUN.

    ...
  • 排序方式：
  • 1
  • /