2. 程式人生 > >「Computer Vision」Note on Interactive Full Image Segmentation

「Computer Vision」Note on Interactive Full Image Segmentation

阿新 • • 發佈：2018-12-14

QQ Group: 428014259
Sina Weibo：小鋒子Shawn
Tencent E-mail：[email protected]
http://blog.csdn.net/dgyuanshaofeng/article/details/84865335

作者：Eirikur Agustsson, Jasper R. R. Uijlings, Vittorio Ferrari
單位：Google Research

Jasper R. R. Uijlings [Google Scholar]，大家可能比較熟悉，就是Selective Search的作者。

0 摘要

解決互動全圖標註任務，該任務的目標是生成所有object和stuff區域的分割圖。屬於scribble方法，即畫粗線、塗鴉。改編Mask R-CNN。在4個extreme clicks和個改正的scribbles每區域，所提出方法可提升5%IoU，接近90%的IoU。

1 介紹

提到標註成本，Cityscapes每一影象需要1.5小時，COCO+stuff每一影象需要19分鐘。提到scribble corrections比click correstions更為自然。

2 相關工作

互動全圖分割 提及Fluid Annotation

3 作者的互動分割模型

主要包含從extreme points生成分割圖，然後利用scribble corrections嵌入Mask R-CNN中，生成校正的分割圖。

3.1 模型結構

改編的Mask R-CNN，由接收單一RGB輸入，變為接收3種輸入資訊。第一種就是RGB影象，第二種就是extreme points和scribble corrections生成的標註圖，假設 S

1 , . . . , S N \text S_{1},...,\text S_{N}
，第三種就是，由extreme points決定的包圍盒box，假設 b 1 , . . . , b N \text b_{1},...,\text b_{N} 。N為標註者想要把圖分割成N個區域。

[1] Interactive Full Image Segmentation 2018 [paper]

相關推薦

「Computer Vision」Note on Interactive Full Image Segmentation

QQ Group: 428014259 Sina Weibo：小鋒子Shawn Tencent E-mail：[email protected] http://blog.csdn.net/dgyuanshaofeng/article/details/84865335 作者：

「Computer Vision」Note on Attention-guided Unified Network

QQ Group: 428014259 Sina Weibo：小鋒子Shawn Tencent E-mail：[email protected] http://blog.csdn.net/dgyuanshaofeng/article/details/84962899 [1]

「Computer Vision」Note on Facial Landmark Machines

QQ Group: 428014259 Sina Weibo：小鋒子Shawn Tencent E-mail：[email protected] http://blog.csdn.net/dgyuanshaofeng/article/details/84962877 [1]

「Computer Vision」Note on Feature Denoising

QQ Group: 428014259 Sina Weibo：小鋒子Shawn Tencent E-mail：[email protected] http://blog.csdn.net/dgyuanshaofeng/article/details/84962796 [1]

「Computer Vision」Note on SlowFast Networks

QQ Group: 428014259 Sina Weibo：小鋒子Shawn Tencent E-mail：[email protected] http://blog.csdn.net/dgyuanshaofeng/article/details/84962772 [1]

「Computer Vision」Note on Scale-aware Multi-level Guidance

QQ Group: 428014259 Sina Weibo：小鋒子Shawn Tencent E-mail：[email protected] http://blog.csdn.net/dgyuanshaofeng/article/details/84946772 [1]

「Computer Vision」Note on Stacked Dense U-Nets

QQ Group: 428014259 Sina Weibo：小鋒子Shawn Tencent E-mail：[email protected] http://blog.csdn.net/dgyuanshaofeng/article/details/84865357 [1]

「Computer Vision」Note on Bag of Tricks for Image Classification

QQ Group: 428014259 Sina Weibo：小鋒子Shawn Tencent E-mail：[email protected] http://blog.csdn.net/dgyuanshaofeng/article/details/84843079 [1]

「Computer Vision」Note on Pseudo-3D Residual Net (P3D ResNet)

QQ Group: 428014259 Sina Weibo：小鋒子Shawn Tencent E-mail：[email protected] http://blog.csdn.net/dgyuanshaofeng/article/details/84996428 [1]

「Computer Vision」Note on Faster Training of Mask R-CNN

QQ Group: 428014259 Sina Weibo：小鋒子Shawn Tencent E-mail：[email protected] http://blog.csdn.net/dgyuanshaofeng/article/details/84995989 [1]

「Computer Vision」Note on Long-Term Feature Bank

QQ Group: 428014259 Sina Weibo：小鋒子Shawn Tencent E-mail：[email protected] http://blog.csdn.net/dgyuanshaofeng/article/details/84995855 [1]

「Computer Vision」Notes on Hypercolumns

QQ Group: 428014259 Sina Weibo：小鋒子Shawn Tencent E-mail：[email protected] http://blog.csdn.net/dgyuanshaofeng/article/details/83904764 閱讀的

「Computer Vision」Notes on Label Correction

QQ Group: 428014259 Sina Weibo：小鋒子Shawn Tencent E-mail：[email protected] http://blog.csdn.net/dgyuanshaofeng/article/details/83904036 [1]

「Computer Vision」Notes on Deep Feature Pyramid Reconfiguration

QQ Group: 428014259 Sina Weibo：小鋒子Shawn Tencent E-mail：[email protected] http://blog.csdn.net/dgyuanshaofeng/article/details/83834348 [1]

「Computer Vision」Notes on Scale Normalization for Image Pyramids

QQ Group: 428014259 Sina Weibo：小鋒子Shawn Tencent E-mail：[email protected] http://blog.csdn.net/dgyuanshaofeng/article/details/83834327 [1]

「Computer Vision」Notes on Multi-Region CNN

QQ Group: 428014259 Sina Weibo：小鋒子Shawn Tencent E-mail：[email protected] http://blog.csdn.net/dgyuanshaofeng/article/details/83834303 [1]

「Computer Vision」Notes on DeepMultiBox

QQ Group: 428014259 Sina Weibo：小鋒子Shawn Tencent E-mail：[email protected] http://blog.csdn.net/dgyuanshaofeng/article/details/83834290 [1]

「Computer Vision」Notes on OverFeat

QQ Group: 428014259 Sina Weibo：小鋒子Shawn Tencent E-mail：[email protected] http://blog.csdn.net/dgyuanshaofeng/article/details/83834273 [1]

「Computer Vision」Notes on Deep Learning for Generic Object Detection

QQ Group: 428014259 Sina Weibo：小鋒子Shawn Tencent E-mail：[email protected] http://blog.csdn.net/dgyuanshaofeng/article/details/83834249 [1]

「Computer Vision」Notes on Fast Neural Architecture Search

[1] Fast Neural Architecture Search of Compact Semantic Segmentation Models via Auxiliary Cells 2018