Unsupervised Grouping of Local Components for Object Segmentation
In this paper, we propose a novel object segmentation method for image understanding. Due to challenges such as variations in object size, orientation, illumination etc. object segmentation is extraordinarily difficult task in the domain of image understanding. It is well-founded concept that a small portion of the pixel set in an image contributes most in image description. Based on this concept, we hypothesize that an image consists of many components or parts each of which represent a small local area in the image and they are very meaningful in visual perception. For object segmentation, we propose spatial segmentation method on such prototypical components of images. Given an image this segmentation method acts as coarse to fine search for object(s) iteratively. The proposed method demonstrate its excellence in localizing objects in various complex backgrounds, multiple objects in a single image even if they have variation in size, orientation, lighting conditions etc. The detection efficiency of our object detector on our self-collected image set which consists of images from six different object categories climbs up to 93% in average.
- There are currently no refbacks.