OUTDOOR SCENE LABELLING WITH LEARNED FEATURES AND REGION CONSISTENCY ACTIVATION
Yandong Li
1,2
, Ferdous Sohel
2
, Mohammed Bennamoun
2
, Hang Lei
1
1
University of Electronic Science and Technology of China (UESTC),
2
The University of Western Australia (UWA)
1. Introduction
Fig. 1. Top Left: Original image. Top
Right: Ground truth labels of the
original image. Bottom Left: Labelled
image using learned features from
ConvNets. Bottom Right: Scene
labelling with the proposed Region
Consistency Activation (RCA).
1.1 Objective of scene
labelling
1.2 Challenges of scene
labelling
1.3 Our main contributions
32×32
64×64
128×128
Multiscale superpixel
representation
(Sec. 3.1)
Ultrametric
Contour Map
(UCM)
Threshold = 0.1 Threshold = 0.3 Threshold = 0.6
Probability of Region Consistency (PRC)
(Sec. 3.4)
Convolution
Pooling
Convolution
Learned
features
Pooling
ConvNets based feature learning
Trained ConvNets features(Sec. 3.2)
General ConvNets features(Sec. 3.3)
Region Consistency
Activation (RCA)
(Sec. 3.5)
Input image
2. Background
2.1 Classic strategies for scene labelling
!
"
!
#
!
$
"%
!
#&'()*+)$",
!
'-.-/
#
!
0-1'
!
1'
2.2 Recent development in scene labelling
%
2%-./
3. Method
Fig. 2. Block diagram of our proposed scene labelling framework
"13
4134'5
3.1 Multiscale superpixel representation
#&
#&
Table. 1. Constructed ConvNets
models
3.2 Trained ConvNets
features
!
Y6"
!
L:
!
l
i
:
!
H
0
6&
!
W67
!
b6$
P( | ;( , ))
i
l
0
Y(i) L H W b
3.3 General ConvNets features
" %*'
!
89:
!
(&
!
7&#;18<5=>-
**'
!
8?*'45<<<
!
*'=4=<<<
!
$0"*'
0
!
1+$@A;
!
4(5
!
#
3.4 Probability of Region
Consistency
A06
7
B6
Fig. 3. The relationship
between PRC and the threshold
to UCM.
log( )PRC threshold
45
48
3.5 Region Consistency Activation
"6
"4)C
6
( )
( ( ))
i
i
i
sum i
Y
P
Y
4=
( , ) ( , ) ( )Y i j Y i j PRC P j
4?
*13
!
'
A0
!
"1
4B8
!
4B
=
!
&
4B?
4. Experiments
Fig. 4. From row 1 to row 4 are original images, ground truth labels,
results based on the general ConvNets features and results based on
ConvNets features plus RCA. Column (a) to (e) are /ve scene labelling
examples from the Stanford Background dataset. Image (e) is an
example which illustrates that RCA may lead to a lower scene labelling
accuracy without the support from accurately classi/ed pixels (it
however achieved better visual consistency).
4.1 Examples on Stanford Background dataset
4.2 Results on Stanford Background and SIFT
Flow datasets
Table. 2. State-of-the-art
methods on Stanford
Background dataset
Table. 3. State-of-the-art
methods on SIFT Flow dataset
5. Conclusions
2-
+
13D
相关内容推荐
锁锁屏杂志杂志穿搭图儿童绘本征稿奇幻小说征稿平台中秋国庆征稿晋江雕塑征稿厅征稿地质科普杂志龚俊杂志满贯杂志实体店建邺教育杂志融媒体征稿小学作文征稿 20172016诗词征稿校园美术展征稿组干类杂志地租日咖杂志淋巴癌杂志东南文化 征稿启事脊梁杂志社幻城记杂志国庆征稿启事崔连准杂志九几年的杂志天风杂志邮箱杨之楹杂志费加罗杂志吴磊杂志西装造型阿衰漫画杂志会议论文征稿信息千玺宝马杂志团契征稿华文月刊征稿人物杂志图文排版征稿投稿方案2017儿童文学征稿美食杂志征稿日记报征稿中国书画征稿启事夔山杂志页沈小婷杂志原创衣服征稿中州学刋杂志国际学生书法征稿2015年书画征稿征稿比赛画龙与豹杂志征稿手绘安全伴我成长征稿时尚杂志股东nanolife征稿亚洲杂志社英国奥秘杂志老科协征稿启示公园书法征稿通知各界征稿聊城文艺征稿窝蜜瓜杂志人民教育征稿武功山征稿致趣杂志馆富裕层的杂志旅游会议征稿扬州原创诗歌征稿美学杂志网站设计征稿许可居家杂志风拍照征稿豆瓣2014年摄影征稿父亲节征稿石油征稿诗词征稿网贾正龙杂志头条有奖征稿枫叶湘潭征稿让粉丝买杂志商丘征稿城投相关杂志党员征稿山东书画征稿exo杂志架疫情学术征稿感应器杂志铜仁杂志社强国征稿杨幂新杂志烙画征稿2016年文学奖征稿启事万众一心征稿国学动画征稿篮球报纸杂志初中征稿启事征稿学报正在征稿00后征稿漫画征稿 2022征稿回复制裁杂志社长沙安防杂志2017山东画展征稿中医杂志水平捉妖app征稿ka文库征稿剧本征稿启事校园征稿主题征稿函平台沙地杂志社肖战照片杂志vouge杂志怎么读九号酒馆征稿国际摄影大赛征稿征稿其实杂志铺小花仙诗词最新征稿大赛县征稿下一代征稿农村 征稿杂志素描画自拍征稿哈尔滨啤酒征稿中国文化报 征稿于赛丽杂志双年展征稿版征稿杂志社邮件硝酸钾杂志大班儿童杂志推荐新纪实杂志内容地理杂志美国龚俊杂志2021文艺征稿通知微电机征稿楹联征稿罗恩的杂志农发行青年征稿芭比杂志官网西影征稿ok精彩杂志推荐2015征稿比赛持续征稿国外放射杂志杂志封面抖音学校杂志封面图片苏州征稿时装杂志王源ok精彩杂志推荐高考征稿启事冷亮型杂志小说系列征稿2016全国诗歌大赛征稿男生女生征稿疫情征稿主题征稿启事400捉妖app征稿乾陵赋征稿征稿计划书驻村征稿书法研讨会征稿大漫王杂志相声大赛征稿一念杂志校长阅刊杂志全国第二届行书展征稿丽水画展征稿烟有什么杂志读者杂志投稿散文2017山东画展征稿养眼美女私杂志光夜杂志拍摄2017全国书法征稿广东书画征稿菊韵征稿征稿中北京美协最新征稿2019年书画大赛征稿某本杂志英语船海杂志投稿征稿推文引言大诗征稿江苏省油画展征稿2022廉政书画征稿酒文化书画征稿2022动画征稿mk2杂志新传论文征稿雪庐老人杂志杂志拼贴作业图片校庆文艺征稿五年级征稿发放征稿函钎九杂志花絮校友征稿九月 杂志杂志社征稿插画河北省画展征稿诊疗康复杂志中国梦征稿史学论坛征稿2022千叶刀杂志杭州西泠印社征稿
最新资源