1. What are the contributions mentioned in the paper "Correspondence driven saliency transfer" ?
In this paper, the authors show that large annotated data sets have great potential to provide strong priors for saliency estimation rather than merely serving for benchmark evaluations.. To this end, the authors present a novel image saliency detection method called saliency transfer.. The authors then introduce two refinement measures to further refine the saliency maps and apply the random-walk-with-restart by exploring the global saliency structure to estimate the affinity between foreground and background assignments.
read more
2. What is the shortest path to the virtual node?
The second type of saliency distance d2 is defined as the geodesic distance between superpixel ri and virtual node v :d2(ri) = min .w( p, q), p, q ∈ Cri ,v, (10) Cri ,v p,qwhere Cri ,v is a path connecting nodes ri and v . d2 for region ri is computed as the accumulated distance along its shortest path to the virtual node.
read more
3. What is the method for detecting salient objects?
Thanks toour RWR based optimization, their method is able to detectsalient objects accurately despite similar appearance to the background regions.
read more
4. How much time does scene matching take?
Scene matching [22] occupies almost all the computation time, since the GIST descriptor can be pre-stored for an annotated dataset and scene retrieval takes little time.
read more
![Fig. 1. We can ask how to identify correctly the salient region in complex scenario (a). The state-of-the-art methods, e.g., (b) the contrast prior based RC [11] and (c) the background prior based MR [15], face with ambiguity since they have no mechanism to incorporate additional contextual information. Our correspondence-based saliency transfer method (d) utilizes the saliency prior (f) from a set of support images (e) that share similar contextual scene information with the input image.](/figures/fig-1-we-can-ask-how-to-identify-correctly-the-salient-1k4aym25.webp)
![Fig. 6. Comparison of saliency maps with eight state-of-the-art methods. From top to bottom: Input images, ground-truth, saliency maps generated by GS12 [31], SF12 [30], HS13 [35], MC13 [14], MR13 [15], wCtr14 [18], BSCA15 [21], BL15 [36] and our method. Note that the proposed method generates more reasonable saliency maps compared with the state-of-the-art.](/figures/fig-6-comparison-of-saliency-maps-with-eight-state-of-the-3ly9038u.webp)
![Fig. 7. Statistical comparison with 8 alternative saliency detection methods using MSRA-5000 [9], ECCSD [35], DUT-OMRON [15] and PASCAL-S [37] datasets: (a) PR curves, (b) F-measure, (c) MAE.](/figures/fig-7-statistical-comparison-with-8-alternative-saliency-3gfe30hw.webp)
![Fig. 3. Illustration of correspondence-based warping. We choose the N best matching images and patches according to the feature similarity between the warped image and the input image using SIFT descriptors.(a) Input image I . (b) A support image Ii retrieved from the reference dataset via GIST matching. (c) The annotation of Ii . (d) Pixel-wise correspondences between I and Ii established via [22]. (e) Warped image of Ii according to the correspondences in (d), which is similar to test image I . (f) Warped annotation gi of (c) according to the correspondences in (d).](/figures/fig-3-illustration-of-correspondence-based-warping-we-choose-2vw57swp.webp)