Simultaneous semi-supervised segmentation of category-independent objects from a collection of images

Abstract

This work is about simultaneous segmentation of different foreground objects from a collection of images with heterogeneous contents. Our idea is to propagate the segmentation information between images in order to detect foreground objects in all these images simultaneously, under the hypothesis of using categorized or uncategorized images, rather than resorting to image co-segmentation that forces the use of similar categorized images. In fact, given an input image, the objective is to integrate seamlessly other images in the general foreground model, in order to benefit the segmentation of the foreground objects in this image. Indeed, the proposed method aggregates general information, on foregrounds as well as on backgrounds, from a collection of images. To this end, the method is based on an energy minimization function. The linear dependence of the foreground histograms is firstly estimated to optimize the proposed energy function. Then, an iterative optimization of each image permits to remarkably optimize the final segmentation result for all images composing the input collection. Extensive experiments demonstrate that the suggested method allows full-object segmentation of the foreground from a collection of images composed of different classes of objects. Indeed, the validation of the accuracy on four challenging datasets (iCoseg, Oxford Flowers, Caltech101 and Berkeley) shows that the proposed method compares favorably with the state-of-the-art of foreground object segmentation from a collection of images. Besides, it has the challenging ability to deal accurately with uncategorized objects.