Dense ASPP for Semantic Segmentation in Street Scenes

  • Slides: 10
Download presentation
Dense. ASPP for Semantic Segmentation in Street Scenes CVPR 2018 Maoke Yang, Kun Yu,

Dense. ASPP for Semantic Segmentation in Street Scenes CVPR 2018 Maoke Yang, Kun Yu, Chi Zhang*, Zhiwei Li*, Kuiyuan Yang* Deep. Motion. Inc * MSRA Speaker: Zhixuan Li 2019/03/11 Zhixuan Li 1

Important challenge in Scene Segmentation • Scale variation • People • Cars 2019/03/11 Zhixuan

Important challenge in Scene Segmentation • Scale variation • People • Cars 2019/03/11 Zhixuan Li 2

Importance of Large Reception Field Small Object Small Kernel 2019/03/11 Zhixuan Li Big Object

Importance of Large Reception Field Small Object Small Kernel 2019/03/11 Zhixuan Li Big Object Small Kernel Big Object Big Kernel 3

How to enlarge Reception Field • Atrous convolution is proposed to enlarge the size

How to enlarge Reception Field • Atrous convolution is proposed to enlarge the size of reception field and dicrease the cost of computing simultaneously. 2019/03/11 Zhixuan Li 4

How to enlarge Reception Field • ASPP is proposed to concatenate feature maps generated

How to enlarge Reception Field • ASPP is proposed to concatenate feature maps generated by atrous convolution with different dilation rates. 2019/03/11 Zhixuan Li 5

Disadvantage of ASPP • When training high resolution images, the size of reception field

Disadvantage of ASPP • When training high resolution images, the size of reception field will never become big enough. • And large enough dilation ratio (d>24) will decrease the power of model. 2019/03/11 Zhixuan Li 6

Dense. ASPP • Layer cascade • Through cascade, applying d=6 conv on the feature

Dense. ASPP • Layer cascade • Through cascade, applying d=6 conv on the feature map of d=3 conv, we can actually get the bigger receptive field than a normal d=6 conv can get. 2019/03/11 Zhixuan Li 7

Network Structure 2019/03/11 Zhixuan Li 8

Network Structure 2019/03/11 Zhixuan Li 8

Rank of the experiment results 2019/03/11 Zhixuan Li 9

Rank of the experiment results 2019/03/11 Zhixuan Li 9

Conclusion • Cascade is an important thought, separating a big problem to some simple

Conclusion • Cascade is an important thought, separating a big problem to some simple problems is very helpful sometimes. 2019/03/11 Zhixuan Li 10

Semantic segmentation Semantic segmentation using convolutional networks wSemantic segmentation Semantic segmentation using convolutional networks w
Nonparametric Semantic Segmentation Tackgeun You Motivation Semantic SegmentationNonparametric Semantic Segmentation Tackgeun You Motivation Semantic Segmentation
Dense semantic image segmentation with objects and attributesDense semantic image segmentation with objects and attributes
Market Segmentation MARKET SEGMENTATION Market segmentation is basedMarket Segmentation MARKET SEGMENTATION Market segmentation is based
SEGMENTATION Cours de Segmentation PLAN Introduction Segmentation parSEGMENTATION Cours de Segmentation PLAN Introduction Segmentation par
Segmentcia farebnho obrazu Image segmentation Image segmentation SegmentationSegmentcia farebnho obrazu Image segmentation Image segmentation Segmentation
Image Segmentation Chapter 10 Image Segmentation Image segmentationImage Segmentation Chapter 10 Image Segmentation Image segmentation
Image Segmentation Image Segmentation Segmentation divides an imageImage Segmentation Image Segmentation Segmentation divides an image
Image Segmentation 1 Image Segmentation s Image segmentationImage Segmentation 1 Image Segmentation s Image segmentation