《Deep Self-Taught Learning for Weakly Supervised Object Localization》
2017-09-30 18:31
627 查看
本文提出依靠检测器自身不断改进训练样本质量,不断增强检测器性能的一种全新方法,破解弱监督目标检测问题中训练样本质量低的瓶颈。
Weakly Supervised Objection Localization——弱监督目标定位
也就是说我们训练的图像只有针对图像的【注释】,没有包含物体的矩形框,入论文中的图片所示:
可以很显然的看出,这种无矩形框的弱监督学习要比有矩形框的学习 难很多。
本文的算法流程图如图所示:
候选区域提取使用的是Hypothesis-CNN-Pooling (HCP) [26]方法,讲图像中属于同一类物体的候选区域提取出来。
2. Reliabe Seed Proposal Generation
上一个步骤提供了若干个候选区域,我们需要使用图论的知识,选择出一个最好的候选区域。
关键是:将此问题 描述为 dense subgraph discovery (DSD)problem
通过上图,我们可以观察到,最好的候选区域位于graph G 节点网络的中点位置,且他的连接点最多。
坏的候选区域,对于不断提高的 Fast R-CNN。很难获得relative improvement
好的候选区域,对于不断提高的 Fast R-CNN。可以获得relative improvement
Weakly Supervised Objection Localization——弱监督目标定位
也就是说我们训练的图像只有针对图像的【注释】,没有包含物体的矩形框,入论文中的图片所示:
可以很显然的看出,这种无矩形框的弱监督学习要比有矩形框的学习 难很多。
本文的算法流程图如图所示:
3.Deep Self-Taught Learning for WSL
3.1第一部分是Seed Sample Acquisition 初始样本的采集工作;
1. Image-to-Object Transfer候选区域提取使用的是Hypothesis-CNN-Pooling (HCP) [26]方法,讲图像中属于同一类物体的候选区域提取出来。
2. Reliabe Seed Proposal Generation
上一个步骤提供了若干个候选区域,我们需要使用图论的知识,选择出一个最好的候选区域。
关键是:将此问题 描述为 dense subgraph discovery (DSD)problem
通过上图,我们可以观察到,最好的候选区域位于graph G 节点网络的中点位置,且他的连接点最多。
3.2第二部分是Online Supportive Sample Harvesting 在线支持样本采集工作;
这一步是通过Fast R-CNN 进一步得到 higher-quality positive sample。通过观察不同的候选区域的 relative improvement of output CNN scores 来自动找出最好的候选区域。坏的候选区域,对于不断提高的 Fast R-CNN。很难获得relative improvement
好的候选区域,对于不断提高的 Fast R-CNN。可以获得relative improvement
相关文章推荐
- 目标定位--Deep Self-Taught Learning for Weakly Supervised Object Localization
- Deep Self-Taught Learning for Weakly Supervised Object Localization
- 论文笔记《IS object localization for free? Weakly-supervised learning with convolutional neural network》
- Two-Phase Learning for Weakly Supervised Object Localization
- 论文阅读Soft Proposal Networks for weekly-supervised object localization
- [EMNLP2017]Deep Residual Learning for Weakly-Supervised Relation Extraction
- 行人属性“Weakly-supervised Learning of Mid-level Features for Pedestrian Attribute Recognition and Loca”
- 论文研读--Weakly Supervised Object Localization with Progressive Domain Adaptation
- Weakly supervised object recognition with convolutional neural networks 论文解读
- Learning color and locality cues for moving object detection and segmentation方法解析
- Self-Supervised Learning for Stereo Matching with Self-Improving Ability
- 目标检测--DSOD: Learning Deeply Supervised Object Detectors from Scratch
- Spatially Supervised Recurrent Convolutional Neural Networks for Visual Object Tracking
- [IDW-CNN]Learning Object Interactions and Descriptions for Semantic Image Segmentation
- 笔记:Semi-supervised domain adaptation with subspace learning for visual recognition (cvpr15)
- cascade learning 级联学习 learning SURF Cascade for Fast band Accurate Object Detection
- 笔记:Semi-supervised Domain Adaptation with Subspace Learning for Visual Recognition
- 物体检测“DSOD: Learning Deeply Supervised Object Detectors from Scratch”
- Learning Dynamic Siamese Network for Visual Object Tracking 阅读笔记
- Semi-supervised Learning for Network-Based Cardiac MR Image Segmentation文章解读