Saliency detection on videos with scene change

Li, Junling; Fang, Meng; Mao, Jingbo

doi:10.1109/icalip.2014.7009845

Cited by 3 publications

(1 citation statement)

References 6 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…Um mapa de saliências pode ser visto como uma imagem cinza, onde os valores dos pixels representam graus de similaridade com propriedades de um objeto de interesse. Esses mapas são comumente explorados para localizar objetos de interesse em uma imagem ou vídeo [12,29,25,24]. Mapas de saliências podem ser gerados a partir de classificação supervisionada, em que os valores associados são probabilidades (ou grau de certeza) da pertinência dos pixels com o objeto [68,26,40], ou por procedimentos não supervisionados, como por exemplo, por técnicas baseadas em segmentação de superpixels [53,11,48,61,62].…”

Section: Contribuiçõesunclassified

Geração de superpixels pela Floresta Geradora Iterativa usando informação de objeto

Arteaga¹

View full text Add to dashboard Cite

Image segmentation through superpixel methods aims partitioning the image in connected regions (i.e., superpixels) such that the objects of interest can be represented by the union of its superpixels. Such result is crucial for many applications, improving the computational performance and allowing to explore mid-level information of the objects involved in the image analysis. Depending of the algorithm, the superpixel segmentation performance may be proportional to the number of regions generated. However, the lack of information of the objects of interest leads efficient results to be often related to an unnecessary oversegmentation, causing a negative impact over the aforementioned objectives. Given the development of efficient methods for superpixel segmentation -whose objects can be efficaciously represented with a low quantity of superpixels -, this work incorporates object information into the Iterative Spanning Forest (ISF) framework. Like ISF, the resulting framework, named Object-based Iterative Spanning Forest (OISF), is composed of three independent steps: (i) initial seed pixels sampling; (ii) superpixel delineation through the Image Foresting Transform (IFT) restricted to the sampled seeds, for a given connectivity function (i.e., path-cost in an image graph); and (iii) seed recomputation, followed by iterative executions of steps (ii) and (iii) for a better displacement of seeds and, consequently, better superpixel delineation. The object information is obtained through a previously generated saliency map, and it is used for incorporating object information in all three steps of OISF. Results indicates a higher efficacy for lower quantities of superpixels, and flexibility in adapting the framework for different applications. Such results are demonstrated in comparison with many different state-of-the-art methods, including ISF-based ones, for two natural image datasets, and one medical.

show abstract