【ICCV 2015】Active Object Localization with Deep Reinforcement Learning


【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

Active Object Localization with Deep Reinforcement Learning

https://arxiv.org/pdf/1511.06015.pdf

做了什么

该论文将强化学习应用于图像中的目标定位。我们用手机看图片时,会通过放大,滑动屏幕等等操作来定位目标:

【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

该论文让智能体学习类似的操作。

怎么做的

对目标定位进行马尔可夫决策过程建模,使用DQN算法让智能体学习定位策略。

动作

总共九个动作:

(1)八个动作用于定位边框的变换:

这些操作如图2所示,分成四个子集:在水平轴(horizontal)和垂直轴(vertical)上移动边框、改变边框的比例(scale)和修改边框的长宽比( aspect ratio)。
【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

边框由其两个角的像素坐标表示:【ICCV 2015】Active Object Localization with Deep Reinforcement Learning。动作对边框的更改值【ICCV 2015】Active Object Localization with Deep Reinforcement Learning【ICCV 2015】Active Object Localization with Deep Reinforcement Learning与边框的当前大小相关:

【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

其中【ICCV 2015】Active Object Localization with Deep Reinforcement Learning,实验中设置为0.2。

例如,

向右水平移动边框可以将【ICCV 2015】Active Object Localization with Deep Reinforcement Learning【ICCV 2015】Active Object Localization with Deep Reinforcement Learning【ICCV 2015】Active Object Localization with Deep Reinforcement Learning相加;

而减小纵横比可以将【ICCV 2015】Active Object Localization with Deep Reinforcement Learning【ICCV 2015】Active Object Localization with Deep Reinforcement Learning相减,并将【ICCV 2015】Active Object Localization with Deep Reinforcement Learning【ICCV 2015】Active Object Localization with Deep Reinforcement Learning相加。

请注意,图像平面中的原点位于左上角。

(2)一个动作用于结束目标搜索
【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

触发器(trigger)不变换边框,而是用于指示对象已被定位。一旦执行了这个动作,目标搜索将终止,并在初始位置重新启动边框以进行新一轮目标搜索。触发器还对图像进行了修改:它用黑➕标记了上次搜索得到的目标区域,通过这样阻止已定位的目标再次被定位,实现多目标的定位。

【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

一些例子:

【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

状态

【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

状态是一个元组【ICCV 2015】Active Object Localization with Deep Reinforcement Learning,其中【ICCV 2015】Active Object Localization with Deep Reinforcement Learning是边框包围区域的特征向量,【ICCV 2015】Active Object Localization with Deep Reinforcement Learning所采取历史动作组成的向量。

边框内的区域都会被扩展到原始框周围的16个像素,变换为224×224以匹配网络的输入大小,输入预训练CNN,输出4096维的特征向量【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

历史向量【ICCV 2015】Active Object Localization with Deep Reinforcement Learning 包含10个历史动作。每个动作都用one-hot向量表示。这意味着【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

奖励

【ICCV 2015】Active Object Localization with Deep Reinforcement Learning为目前的边框,【ICCV 2015】Active Object Localization with Deep Reinforcement Learning为目标边框。【ICCV 2015】Active Object Localization with Deep Reinforcement Learning【ICCV 2015】Active Object Localization with Deep Reinforcement Learning的IoU为:

【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

八个变换动作的奖励

当智能体选择动作【ICCV 2015】Active Object Localization with Deep Reinforcement Learning从状态【ICCV 2015】Active Object Localization with Deep Reinforcement Learning移动到状态【ICCV 2015】Active Object Localization with Deep Reinforcement Learning时,边框由【ICCV 2015】Active Object Localization with Deep Reinforcement Learning变为【ICCV 2015】Active Object Localization with Deep Reinforcement Learning。智能体获得的奖励:
【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

直观地说,等式(2)表示如果从状态【ICCV 2015】Active Object Localization with Deep Reinforcement Learning到状态【ICCV 2015】Active Object Localization with Deep Reinforcement Learning时IoU变大,则奖励为【ICCV 2015】Active Object Localization with Deep Reinforcement Learning,否则为【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

终止动作的奖励

【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

其中【ICCV 2015】Active Object Localization with Deep Reinforcement Learning是终止动作,【ICCV 2015】Active Object Localization with Deep Reinforcement Learning是终止奖励,实验中设置为3.0,【ICCV 2015】Active Object Localization with Deep Reinforcement Learning是一个阈值,表示允许将检测到区域视为真正(TP)的最小IoU,实验中【ICCV 2015】Active Object Localization with Deep Reinforcement Learning设置为0.6。

结果

所有参与区域(All joining regions, AAR):对智能体处理的所有区域进行评分。

终端区域(Terminal regions,TR):只考虑智能体使用触发器指示为存在目标的区域。

准确率评估

【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

召回率评估

【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

图5绘制了正确检测到对象所需的步数的分布图。

【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

定性评估
【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

错误例子
【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

敏感性分析

评估特征有:遮挡(occ)、截断(trn)、大小、纵横比(asp)、物体视点(view)和部分可见(parts)。

【ICCV 2015】Active Object Localization with Deep Reinforcement Learning

文章出处登录后可见!

已经登录?立即刷新

共计人评分,平均

到目前为止还没有投票!成为第一位评论此文章。

(0)
xiaoxingxing的头像xiaoxingxing管理团队
上一篇 2022年5月30日 上午11:19
下一篇 2022年5月30日

相关推荐