文章来源:CVPR 2016
下载链接:https://arxiv.org/abs/1603.04530
解决的问题:
① Previous algorithms efforts lift edge detection to a higher abstract level, but still fall below human perception due to their lack of object-level knowledge. Different from previous low-level edge detection, our algorithm focuses on detecting higher-level object contours.
② However, since it is very challenging to collect high-quality contour annotations, the available datasets for training contour detectors are actually very limited and in small scale.
文章内容:
In this paper, we scale up the training set of deep learning based contour detection to more than 10k images on PASCAL VOC . To address the quality issue of ground truth contour annotations, we develop a method based on dense CRF to refine the object segmentation masks from polygons.
受到完全卷积网络和解卷积网络在语义分割上成功应用的启发,作者开发了一个完全卷积编码-解码网络(CEDN)用于提取目标轮廓,该网络结构图如图1所示。该网络在训练时,作者固定编码器的参数(vgg-16),而只优化解码器的参数。这保持了编码器的泛化能力,并使解码器网络可以容易地与其他任务进行组合。
目前有大量的算法用于生成边界框或分段对象提案,但是他们的边界框通常不能提供准确的对象定位。为了解决这一问题,作者开发了一种基于dense CRF的方法来从多边形中细化对象分割蒙板。
网络结构:
文章的网络结构如图1所示,其在VGG-16 net 上进行了改进。
编码:VGG-16 net
解码:反池化-卷积-激活-dropout
详细内容可参考:CEDN网络结构
网络的 encoder 部分直接使用 VGG模型对应的部分,然后使用 decoder还原尺寸大小。其中解码器网络设置如表1所示,损失函数采用simply the pixel-wise logistic loss。
文章特点:
与以前的low-level边缘检测不同,该算法侧重于检测higher-level的目标轮廓。另外全卷积的Encoder-Decoder网络,可以给出更higher-level的轮廓检测,准确率大幅超越传统算法;以此为基础可产生高质量的segmented object proposals。
存在的不足:
本文方法的F-score = 0.57和上限的F-score = 0.74之间仍然存在很大的性能差距,还需要进一步的改进研究。