权益分类	功能权益	普通用户	{{item.name}}会员
{{category.name}}	{{benefitItem.name}}

Incremental Semantic Segmentation Learning

增量语义分割学习

基本信息

批准号：
21J13152
负责人：
ZHANG KAIPENG
金额：
$ 0.96万
依托单位：
The University of Tokyo
依托单位国家：
日本
项目类别：
Grant-in-Aid for JSPS Fellows
财政年份：
2021
资助国家：
日本
起止时间：
2021-04-28 至 2023-03-31
项目状态：
已结题

来源：
https://kaken.nii.ac.jp/en/grant/KAKENHI-PROJECT-21J13152/
关键词：
Semantic segmentation Neural network Few-shot learning Active learning Neural routing by memory

项目摘要

This research aims to improve semantic segmentation through two solutions, including the passive and active solutions. The passive solution provides the computer with a few new annotated data for new categories to make the computer able to segment the image regions of new categories. The active solution makes the computer able to discover valuable data during running and use them to improve the model.In 2021, we completed the passive solution by proposing a method named Segmentation by Dynamic Prototype (SDP). SDP does segmentation by searching each pixel's features nearest prototype in feature space. A prototype is a representative feature of a class. During running, it is dynamically constructed by a few new annotated data and old data. We submitted this work to a journal, and it is under review so far.As for the active solution, we proposed a continual active learning method for semantic segmentation. It can continually select informative images to annotate and feed them to the model to improve accuracies. But the improvement is not satisfactory so far, and we will do more research in the next.Besides, during the research, we found large redundant storage and RAM resources in cloud servers. Thus, we proposed a method named Neural Routing by Memory, which utilizes the redundant resources to improve accuracies. The work was accepted by NeurIPS 2021.

本研究旨在通过两种解决方案来改进语义分割，包括被动方案和主动方案。被动解决方案向计算机提供新类别的一些新的注释数据，以使计算机能够分割新类别的图像区域。主动解使得计算机能够在运行过程中发现有价值的数据，并利用这些数据来改进模型。2021年，我们提出了一种动态原型分割(SDP)的方法来完成被动解。SDP通过在特征空间中搜索与原型最近的每个像素的特征来进行分割。原型是类的代表性特征。在运行过程中，它是由几个新的注释数据和旧的数据动态构建的。我们将这项工作提交给了一家期刊，目前正在进行审查。对于主动解决方案，我们提出了一种连续主动学习的语义切分方法。它可以不断地选择信息丰富的图像进行注释，并将其提供给模型以提高准确性。但到目前为止，改进效果并不理想，我们将在下一步做更多的研究，此外，在研究过程中，我们发现云服务器中存在着大量的冗余存储和RAM资源。因此，我们提出了一种基于记忆的神经网络路由方法，该方法利用冗余资源来提高精度。这项工作被NeurIPS 2021接受。