人工智能培训

搜索

人工智能论文:CODA:将对象计数到尺度感知的对抗密度适应(CODA: Counting Objects via Scale-aware Adversarial

[复制链接]
luolei 发表于 2019-3-26 11:56:11 | 显示全部楼层 |阅读模式
luolei 2019-3-26 11:56:11 388 0 显示全部楼层
人工智能论文:CODA:将对象计数到尺度感知的对抗密度适应(CODA: Counting Objects via Scale-aware Adversarial Density Adaption)人群计数的最新进展已经在越来越复杂的卷积神经网络设计中取得了有希望的结果。然而,由于不可预测的域转移,将训练模型推广到看不见的场景是不理想的。受到不同情景密度图共享相似局部结构的观察的启发,我们在本文中提出了一种新颖的对抗学习方法,即CODA(\ emph {Counting Objects viascale-aware adversarial Density Adaption})。为了处理不同的对象尺度和密度分布,我们使用来自源域和目标域的多尺度pyramidpatches进行对抗性训练。除了跨金字塔输入级别的限制约束外,还可以针对不同比例生成一致的对象计数。大量实验表明,与未经计数的自适应模型相比,我们的网络在看不见的数据集上产生了更好的结果。值得注意的是,我们的CODA的性能可与目标数据集中训练有素的最先进的全监督模型相媲美。进一步的分析表明,我们的密度自适应框架可以毫不费力地扩展到具有不同对象的场景。\ emph {此代码可通过此https URL获得。}
Recent advances in crowd counting have achieved promising results withincreasingly complex convolutional neural network designs.However, due to theunpredictable domain shift, generalizing trained model to unseen scenarios isoften suboptimal.Inspired by the observation that density maps of differentscenarios share similar local structures, we propose a novel adversariallearning approach in this paper, i.e., CODA (\emph{Counting Objects viascale-aware adversarial Density Adaption}).To deal with different objectscales and density distributions, we perform adversarial training with pyramidpatches of multi-scales from both source- and target-domain.Along with aranking constraint across levels of the pyramid input, consistent object countscan be produced for different scales.Extensive experiments demonstrate thatour network produces much better results on unseen datasets compared withexisting counting adaption models.Notably, the performance of our CODA iscomparable with the state-of-the-art fully-supervised models that are trainedon the target dataset.Further analysis indicates that our density adaptionframework can effortlessly extend to scenarios with different objects.\emph{The code is available at this https URL.}人工智能论文:CODA:将对象计数到尺度感知的对抗密度适应(CODA: Counting Objects via Scale-aware Adversarial Density Adaption) JZhzzdj0mrUze66M.jpg
URL地址:https://arxiv.org/abs/1903.10442     ----pdf下载地址:https://arxiv.org/pdf/1903.10442    ----人工智能论文:CODA:将对象计数到尺度感知的对抗密度适应(CODA: Counting Objects via Scale-aware Adversarial Density Adaption)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

luolei当前离线
新手上路

查看:388 | 回复:0

快速回复 返回顶部 返回列表