Predicting Segmentation "Easiness" from the Consistency for Weakly-Supervised Segmentation
Predicting Segmentation "Easiness" from the Consistency for Weakly-Supervised Segmentation
复制标题
DOI：
10.1109/acpr.2017.124
复制
发表时间：
2017-11
期刊：
影响因子：
--
通讯作者：
Wataru Shimoda;Keiji Yanai
中科院分区：
文献类型：
--
作者：
Wataru Shimoda;Keiji Yanai
弱监督分割已经引起了很多关注，因为它的成本非常高，以创建全监督分割的逐像素标注的图像数据集。最近，通过Wei等人提出的具有粗略估计的初始掩码的全监督分割模型的（重新）训练方法取得了很大进展。[25]。然而，初始估计的掩模往往包括一些噪声，这有时会导致错误的结果。因此，在本文中，我们专注于提高初始估计掩码的质量，以（重新）训练全监督分割模型。我们提出了一种新的算法来检索“好种子”，通过预测分割的“容易”的图像之间的一致性输出与不同的条件。我们证明了训练数据质量和所选图像数量之间存在权衡，并且我们提出的方法可以使用数据增强来改进训练模型。我们已经在Pascal VOC 2012分割基准数据集上的弱监督分割设置中实现了最先进的技术。
Weakly-supervised segmentation has come to draw a lot of attention, since it costs very high to create pixel-wise an-notated image datasets for fully-supervised segmentation. Recently, it has achived great progress by the method of (re-)training of a fully-supervised segmentation model with roughly estimated inital masks which is proposed by Wei et al. [25]. However, the initial estimated masks tend to include some noise, which sometimes causes erroneous results. Therefore in this paper we focus on improving of quality of initial estimated masks for (re-)training of a fully-supervised segmentation model. We propose a novel algorithm to retrieve "good seeds" by predicting segmentation "Easiness" of images based on consistency among the out-puts with different conditions. We show that there is a trade-off between training data quality and the number of selected images, and our proposed method can improved the trained model using data augmentation. We have achieved state-of-the-art in a weakly-supervised segmentation setting on Pascal VOC 2012 segmentation benchmark dataset.