PROJECT TITLE :

Random Geometric Prior Forest for Multiclass Object Segmentation - 2015

ABSTRACT:

Recent advances in object detection have led to the event of segmentation by detection approaches that integrate high-down geometric priors for multiclass object segmentation. A key yet beneath-addressed issue in utilizing high-down cues for the matter of multiclass object segmentation by detection is efficiently generating robust and correct geometric priors. In this paper, we have a tendency to propose a random geometric previous forest theme to get object-adaptive geometric priors efficiently and robustly. Within the theme, a testing object initial searches for training neighbors with similar geometries using the random geometric prior forest, and then the geometry of the testing object is reconstructed by linearly combining the geometries of its neighbors. Our theme enjoys several favorable properties compared with conventional ways. 1st, it is strong and terribly quick as a result of its inference does not suffer from unhealthy initializations, poor local minimums or advanced optimization. Second, the figure/ground geometries of coaching samples are utilized in an exceedingly multitask manner. Third, our theme is object-adaptive however does not need the labeling of parts or poselets, and therefore, it's quite simple to implement. To demonstrate the effectiveness of the proposed scheme, we have a tendency to integrate the obtained high-down geometric priors with conventional bottom-up color cues within the frame of graph cut. The proposed random geometric previous forest achieves the most effective segmentation results of all of the methods tested on VOC2010/2012 and is 90 times faster than the current state-of-the-art methodology.