GM-W3: Two Methods in Saliency Detection
Hongyang Li (oø )
IIAU Lab, Dalian University of Technology [email protected]
Supervisor: Prof. Huchuan Lu (© © © A A A Ç Ç Ç)
September 17, 2013
Hongyang Li (oø ) Two Methods in Saliency Detection
1 Graph-based Manifold Ranking Model Construction
Results
2 Low Rank Matrix Recovery Model Construction Results
3 Summary
Chuan Yang et al.
Saliency Detection via Graph-Based Manifold Ranking CVPR 2013.
Xiaohui Shen and Ying Wu
A Unified Approach to Salient Object Detection via Low Rank Matrix Recovery
CVPR 2012. Oral Presentation.
Graph-based Manifold Ranking
Inspired by PageRank and spectral clustering algorithms, the optimization is:
Steps:
1
superpixels 1 , affinity matrix W, degree matrix D.
2
ranking function, f ∗ = (D − αW) −1 y.
3
background queries, S t , S b , S l , S r ; saliency map S bq .
4
foreground queries, S fq = ¯f ∗ (i ).
1
Website
Hongyang Li (oø ) Two Methods in Saliency Detection
Geodesic distance constraints (Fig.3).
Wise choice of queries (Fig.4, Fig.5).
A two-stage mechanism (Fig.5, Fig.6).
Some Issues
How to find a pervasive parameter?
“These two parameters are empirically chosen, for all the experiments.”
Figure: σ
2= 0.1, α = 0.99
Hongyang Li (oø ) Two Methods in Saliency Detection
How to find a pervasive parameter?
“These two parameters are empirically chosen, for all the experiments.”
Figure: σ
2= 0.2, α = 0.99
Some Issues(Cont’)
How to find a pervasive parameter?
“These two parameters are empirically chosen, for all the experiments.”
Figure: σ
2= 0.1, α = 0.8
Hongyang Li (oø ) Two Methods in Saliency Detection
Some Issues(Cont’)
Figure: Carefully chosen images.
Figure: Stage 1.
Figure: Stage 2.
Hongyang Li (oø ) Two Methods in Saliency Detection
Under some circumstances, the proposed algorithm is confused
with the background.
Low Rank Matrix Recovery for Saliency
Slide Credit: X.Shen.
Hongyang Li (oø ) Two Methods in Saliency Detection
Input 2 : a feature matrix F ∈ R N×D , of the image.
Output: an optimal low-rank matrix L ∗ , of the background; an optimal sparse-noise matrix S ∗ , of the salient object.
Basic Model
(L ∗ , S ∗ ) = arg min L,S (kLk ∗ + λkSk 1 ) s.t. F = L + S
Notation: k · k
∗, nuclear norm; k · k
1, l
1-norm.
Method Solution
3robust PCA.
Saliency Measure kS
ik
1.
Image Decomposition mean-shift clustering.
Feature Representation color, edge, texture; altogether D = 53 features.
2From: J.Yan et al, Visual saliency detection via sparsity pursuit. ISPL,2010.
3From: J.Wright et al, Robust principal component analysis. NIPS,2009.
What is a good way to decompose the image?
Slide Credit: X.Shen.
Hongyang Li (oø ) Two Methods in Saliency Detection
Slide Credit: X.Shen.
Feature Representation
Sparse coding? A sparse encoded vector for each image patch, NOT the sparsity of the saliency over the entire image.
Sln-1: learn a transformation from a set of training images.
Sln-2: higher-level prior integration based on human perception.
Unified Model
(L ∗ , S ∗ ) = arg min L,S (kLk ∗ + λkSk 1 ) s.t. TFP = L + S
T: a learned feature transformation, i.e., linear metric T ∈ R D×D . P: a higher-level prior map, i.e., P = diag (p 1 , p 2 , . . . , p N ), where p i indicates the probability of being salient for the i-th segment.
Hongyang Li (oø ) Two Methods in Saliency Detection
Learn a Feature Transformation(Cont’)
Objective Function
T ∗ = arg min T O(T) , K 1 P K
k=1 kTF k Q k k ∗ − γkTk ∗
s.t. kTk 2 = c where F k and Q k are the feature representation and saliency indicator of the k-th training image respectively.
Elaborations on the function
Method: Gradient Descent, T t+1 = T t − α ∂O(T) ∂T More details, please see paper. However,...
Hongyang Li (oø ) Two Methods in Saliency Detection
Higher-level Prior Integration(Cont’)
How it works? (L
∗, S
∗) = arg min
L,S(kLk
∗+ λkSk
1) s.t. TFP = L + S
More likely to be considered as outliers.
The error terms in S are magnified.
Robust to some of the incorrect higher-level guidance.
Hongyang Li (oø ) Two Methods in Saliency Detection
Figure: Carefully chosen images.
Figure: Northwestern, no semantic prior.
Figure: DLUT, stage 2.
Popular Methods. As of June, 2012
Slide Credit: X.Shen.
Hongyang Li (oø ) Two Methods in Saliency Detection