PROJECT TITLE :

Density-based region search with arbitrary shape for object localisation

ABSTRACT:

Region search is widely used for object localisation in pc vision. When projecting the score of a picture classifier into an image plane, region search aims to search out regions that exactly localise desired objects. The recently proposed region search ways, such as economical subwindow search and efficient region search, typically find regions with maximal score. For some classifiers and situations, the projected scores are nearly all positive or terribly noisy, then maximising the score of a district results in localising nearly the whole images as objects, or causes localisation results unstable. In this study, the authors observe that the projected scores with massive magnitudes are mainly targeted on or around objects. On the basis of this observation, they propose a region search method for object localisation, named level set most-weight connected subgraph (LS-MWCS). It localises objects by looking out regions by graph mode-seeking instead of the maximal score. The score density by localised region can be controlled by a parameter flexibly. They additionally prove an interesting property of the proposed LS-MWCS, which guarantees that the region with desired density can be found. Moreover, the LS-MWCS will be efficiently solved by the assumption propagation theme. The effectiveness of the author's method is validated on the problem of weakly-supervised object localisation. Quantitative results on synthetic and real information demonstrate the superiorities of their methodology compared to different state-of-the-art methods.