目标定位--Deep Self-Taught Learning for Weakly Supervised Object Localization

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
目标定位--Deep Self-Taught Learning for Weakly Supervised Object Localization_第1张图片

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:
目标定位--Deep Self-Taught Learning for Weakly Supervised Object Localization_第2张图片

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。

目标定位--Deep Self-Taught Learning for Weakly Supervised Object Localization_第3张图片

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
目标定位--Deep Self-Taught Learning for Weakly Supervised Object Localization_第4张图片
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。

目标定位--Deep Self-Taught Learning for Weakly Supervised Object Localization_第5张图片

目标定位--Deep Self-Taught Learning for Weakly Supervised Object Localization_第6张图片
目标定位--Deep Self-Taught Learning for Weakly Supervised Object Localization_第7张图片
目标定位--Deep Self-Taught Learning for Weakly Supervised Object Localization_第8张图片

目标定位--Deep Self-Taught Learning for Weakly Supervised Object Localization_第9张图片

你可能感兴趣的:(CVPR2017,目标检测)