INTRODUCTION PROPOSED METHOD Tensor-based High-order Semantic Relation Transfer for Semantic Scene Segmentation

Tensor-based High-order Semantic Relation Transfer for Semantic Scene Segmentation
Heesoo Myeong and Kyoung Mu Lee
Department of ECE, ASRI, Seoul National University, Seoul, Korea
http://cv.snu.ac.kr
INTRODUCTION
Objective Function
PROPOSED METHOD
 We separately deal with the semantic relations transfer problem with respect to Y
Overview
Goal
Results on Jain et al. Dataset
sky
 The quadratic objective function with respect to Y as
 Semantic scene segmentation: identifying and segmenting all objects in a scene
N
sky
N
1
2
2
F (X ) 
w
(
x

x
)


(
x

y
)
,


ijk ,lmn
ijk
lmn
ijk
ijk
2 i , j , k ,l , m , n
i , j ,k
person
tree
building
tree
tree
road
tree
building
tree
building
building
car
tree
person
road
Test image
sky
tree
tree
building
Semantic scene segmentation
…
Query image
Key idea
car
road
car
road
road
Integrated high-order relationSemantic segmentation
sky
sky
building
sky
building
car
person
Annotations of
retrieved images
Predicted top scored
high-order relation
road
sidewalk
Groundtruth
1. For a test image, retrieve similar training images using global features
road
car
car
High-order relations in the training dataset
Previous works & Limitations
 Conventional context models mainly focus on learning pairwise relationships
between objects
 Pairwise relations are not enough to represent high-level contextual knowledge
within images
building
building
2. Apply semantic relation transfer algorithm to transfer third-order semantic relation
from each training image to the test image
3. Integrate the high-order score into MRF optimization framework and obtain
semantic scene segmentation
mountain
𝐸 𝐷 𝑐𝑖 +
𝒥 c =
𝑖
𝐸 𝑃 (𝑐𝑖 , 𝑐𝑗 ) +
𝑖,𝑗
[X
]ijk  xijk
car
𝐸 𝐻 (𝑐𝑖 , 𝑐𝑗 , 𝑐𝑘 )
sidewalk
EXPERIMENTS
Quantitative Results on Standard Datasets
 SIFT Flow dataset (Liu et al., CVPR09):

car
car
Pairwise semantic relation
Our Contributions
 To describe observed third-order semantic relation within the retrieved image, we
define another 𝐾 3 number of semantic tensors Y  {Y111 , Y112 ,..., Y KKK }

yijk
1 if G ( si )  c , G ( s j )  c , G ( sk )  c , ( si , s j , sk )  S retrieved
 
,
0
otherwis
e

 The use of high-order semantic relations for semantic segmentation
 where 𝐺(𝑠𝑖 ) denotes the ground truth class of region 𝑠𝑖
 A novel tensor-based representation of high-order semantic relations
Now, the semantic relation transfer problem is reformulated as the problem of
𝛼𝛽𝛾
estimating the magnitude of confidence scores 𝑥𝑖𝑗𝑘 for all superpixel triplets (𝑠𝑖 , 𝑠𝑗 , 𝑠𝑘 )
and for all object class triplets (𝑐𝛼 , 𝑐𝛽 , 𝑐𝛾 ) based on Y
 A quadratic objective function for learning the semantic tensor and an efficient
approximate algorithm
tree
bison
grass
grass
Ground truth
Proposed
sky
sky
car
mountain
sidewalk
sky
building
sky
building
window
Query
mountain
window
door
door
Ground truth
Proposed
building
window
Query
Ground truth
Proposed
Results on Polo Dataset
 250 training images, 100 test images, 19 labels
person
person
 2,488 training images, 200 test images, 33 labels
horse
grass
 Polo dataset (Zhang et al., CVPR11):
 80 training images, 237 test images, 6 labels
person
horse
Table 1: Per-pixel classification rates and (average per-class rates)
𝛼𝛽𝛾
 The variable 𝑥𝑖𝑗𝑘 indicates confidence score of how likely the region triplet
(𝑠𝑖 , 𝑠𝑗 , 𝑠𝑘 ) would be labeled as (𝑐𝛼 , 𝑐𝛽 , 𝑐𝛾 ), respectively
road sidewalk
building
building
Problem Statement

Query
Proposed
sky
building
𝑖,𝑗
𝐸 𝑃 (𝑐𝑖 , 𝑐𝑗 )
 Jain et al. dataset (Jain et al., ECCV10):
Among superpixels 𝑆 =
(𝑁 is the number of total superpixels, 𝐾 is
the number of object classes), third-order semantic relation is defined as a 𝐾 3 number
of 𝑁 × 𝑁 × 𝑁 semantic tensors X  {X 111 , X 112 ,..., X KKK }
Ground truth
sky
𝐸 𝐻 (𝑐𝑖 , 𝑐𝑗 , 𝑐𝑘 ) ,
where
represents data term,
represents pairwise term, and
is confidence score by the semantic relation transfer algorithm
𝐸 𝐷 (𝑐𝑖 )
road
building
car
Results on LMO Dataset
Semantic Relation Transfer Algorithm
{𝑆 𝑡𝑒𝑠𝑡 , 𝑆 𝑟𝑒𝑡𝑟𝑖𝑒𝑣𝑒𝑑 }
Query
 Use fully connected third-order Markov random field (MRF) model:
car
sky
road
Inference
building
Retrieved images
tree
where 𝑤𝑖𝑗𝑘,𝑙𝑚𝑛 is the triplet-wise similarity between two region triplets (𝑠𝑖 , 𝑠𝑗 , 𝑠𝑘 ) and
(𝑠𝑙 , 𝑠𝑚 , 𝑠𝑛 )
tree
 Exploiting high-order(mostly third-order) semantic relation
sky
car
bison
sky
sky
tree
building
road
car
road
sky
sky
building
building
car
road
tree
sky
grass
Query
Ground truth
horse
grass
grass
person
grass
person
person
horse
horse
grass
horse
grass
Proposed
horse
horse
grass
Query
Ground truth
Proposed
CONCLUSION
 We have presented a novel approach to learn high-order semantic relations of
regions in a nonparametric manner
 We develop a novel semantic tensor representation of the high-order semantic
relations
 We cast the high-order semantic relation transfer problem as a quadratic objective
function of semantic tensors and propose an efficient approximate algorithm