Salient Object Detection for Searched Web Images via Global Saliency

1. Salient Object Detection for Searched Web Images via Global Saliency Peng Wang1Jingdong Wang2 Gang Zeng1Jie Feng1 Hongbin Zha1Shipeng Li2 1Key Laboratory on Machine Perception, Peking University 2Microsoft Research Asia

2. Outline • Introduction • Web Image Database • Proposed Algorithm • Experiments • Conclusion and Discussion

3. Introduction • Want to detect the existence and the location of salient objects for thumbnail images. • Use a learning approach, random forest for solution • Sliding window-based method • Segmentation-based method

4. Introduction

5. Framework

6. Web Image Database • Searched 1100 queries and downloaded 400 thumbnail images for each query. • Each image I(x) is assigned a corresponding label vector y = (o, t, l, b, r). • Bounding box distribution

7. Web Image Database

8. Proposed Algorithm • Input & output space • training features • output labels • mapping function

9. Proposed Algorithm • Creating features

10. Proposed Algorithm • Creating features • Multi-scale Contract (MC) • Center-Surrounding Histogram (CSH) • Region-based Contrast (RC) • Color Spatial Distribution (CSD) With a number of saliency maps Sk(x) in which k = 1,…,K.

11. Proposed Algorithm • Creating features • Combining strategies • Stack Partition each Sk(x) into N = p × p blocks in a grid layout, then extract the average value in each block.

12. Proposed Algorithm • Creating features • Combining strategies • SumUp Apply a non-linear combination of multiple saliency values as: λk : the weight of the kth saliency map

13. Proposed Algorithm • Detecting object existence via classification • Feed the features into the random forest classifier to learn the mapping gc. • Down-sampling the majority class.

14. Proposed Algorithm • Translation and scale invariance feature • Perform a rectification on the saliency map Sall(x). • Fit the two dimensional Gaussian function • Find A, µ,Σ by minimizing the objective • Translate the image center to the position µ = (µx, µy)T • the range of coordinate x: [µx − λσx, µx + λσx]

15. Proposed Algorithm

16. Proposed Algorithm • Localizing object via regression • Learning the posterior distribution p(w|f ) through regression. • Construct the multiple partition {Pz}Zz=1 • Combined through averaging • estimate the position

17. Experiments • Databases • MSRA image set B with images resized into 130 × 130 • Web image database

18. Experiments • Classification evaluation

19. Experiments • Regression evaluation • Training parameters • Number of training image • Minimum size of each node

20. Experiments • Regression evaluation • Effectiveness of features

21. Experiments • Comparison with other approaches • use two measurements • region-based measurement Overlap score

22. Experiments • Comparison with other approaches • use two measurements • edge-based measurement Bounding box Boundary Displacement Error (BDE)

23. Experiments • Comparison with other approaches

24. Experiments • Comparison with other approaches

25. Experiments • Comparison with other approaches

28. Conclusion and Discussion • Presented a large labeled web image database. • A supervised scheme that judges the existence of and predicts the location of the salient object in thumbnail images. • Exploits random forest and global saliency with features created from the saliency maps combining information of multiple channels.