目标定位--Deep Self-Taught Learning for Weakly Supervised Object Localization
2017-04-25 10:15
483 查看
Deep Self-Taught Learning for Weakly Supervised Object Localization
CVPR 2017
https://arxiv.org/abs/1704.05188
什么是 Weakly Supervised Object Localization,refers to learning to localize objects within images with only image-level annotations that simply indicate the presence of an object category.
也就是说我们的训练图像只有 针对图像的 注释,没有包含物体的矩形框,如下图所示
很明显这种弱监督学习要比有矩形框的学习难的多。目前的深度学习也应用到了这个问题。
本文的算法流程如下图所示:
3.1. Seed Sample Acquisition
3.1.1 Image-to-Object Transfer
这里的候选区域提取,我们使用了 文献【26】中的 Hypothesis-CNN-Pooling (HCP) 方法,
将图像中属于同一类物体的候选区域提取出来。
3.1.2 Reliable Seed Proposal Generation
有了若干候选区域,我们需要从中选择出一个最好的候选区域,这里我们使用了图论知识,将其表示为 dense subgraph discovery (DSD) problem
我们可以观察到,最好的候选区域位于同一个物体组成的节点网络的中心点位置。它的连接点最多。
3.2. Online Supportive Sample Harvesting
这一步是通过 Fast R-CNN 进一步得到 higher-quality positive samples。通过观察不同候选区域的 relative improvement (RI) of output CNN scores来自动找出最好的候选区域。
Bad seed samples hardly obtain RI from the increasing detection ability of Fast R-CNN during training. In contrast, high-quality positive samples not selected as seeds mostly gain RI due to the improved detection ability of the model. Therefore, RI is a reliable metric for identifying high-quality positive samples
坏的候选区域 对于 不断提高的Fast R-CNN ,其很难获得 RI
但是好的候选区域 对于 不断提高的Fast R-CNN, 可以得到 RI。
CVPR 2017
https://arxiv.org/abs/1704.05188
什么是 Weakly Supervised Object Localization,refers to learning to localize objects within images with only image-level annotations that simply indicate the presence of an object category.
也就是说我们的训练图像只有 针对图像的 注释,没有包含物体的矩形框,如下图所示
很明显这种弱监督学习要比有矩形框的学习难的多。目前的深度学习也应用到了这个问题。
本文的算法流程如下图所示:
3.1. Seed Sample Acquisition
3.1.1 Image-to-Object Transfer
这里的候选区域提取,我们使用了 文献【26】中的 Hypothesis-CNN-Pooling (HCP) 方法,
将图像中属于同一类物体的候选区域提取出来。
3.1.2 Reliable Seed Proposal Generation
有了若干候选区域,我们需要从中选择出一个最好的候选区域,这里我们使用了图论知识,将其表示为 dense subgraph discovery (DSD) problem
我们可以观察到,最好的候选区域位于同一个物体组成的节点网络的中心点位置。它的连接点最多。
3.2. Online Supportive Sample Harvesting
这一步是通过 Fast R-CNN 进一步得到 higher-quality positive samples。通过观察不同候选区域的 relative improvement (RI) of output CNN scores来自动找出最好的候选区域。
Bad seed samples hardly obtain RI from the increasing detection ability of Fast R-CNN during training. In contrast, high-quality positive samples not selected as seeds mostly gain RI due to the improved detection ability of the model. Therefore, RI is a reliable metric for identifying high-quality positive samples
坏的候选区域 对于 不断提高的Fast R-CNN ,其很难获得 RI
但是好的候选区域 对于 不断提高的Fast R-CNN, 可以得到 RI。
相关文章推荐
- 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
- Deep Auxiliary Learning for Visual Localization and Odometry 基于深度辅助学习的视觉定位和里程计
- [EMNLP2017]Deep Residual Learning for Weakly-Supervised Relation Extraction
- 目标检测--DSOD: Learning Deeply Supervised Object Detectors from Scratch
- Self-Supervised Learning for Stereo Matching with Self-Improving Ability
- deep learning for object detection
- Deep learning-------------Self-Taught Learning to Deep Networks
- 目标检测--A Unified Multi-scale Deep Convolutional Neural Network for Fast Object Detection
- (六)6.12 Neurons Networks from self-taught learning to deep network
- 论文笔记之:Active Object Localization with Deep Reinforcement Learning
- 论文笔记 | Learning Deep Features for Discriminative Localization
- Deep Learning by Andrew Ng --- self-taught
- MIT 深度学习与自动驾驶公开课 Deep Learning for Self-Driving Cars 讲义梗概
- Learning Deep Features for Discriminative Localization
- Modelling Uncertainty in Deep Learning for Relocaliz基于不确定性建模的DL相机重定位
- 【深度学习】深度学习中监督优化入门(A Primer on Supervised Optimization for Deep Learning)
- 目标检测--PVANET: Deep but Lightweight Neural Networks for Real-time Object Detection