Multi-scale Image Co-segmentation

This paper focuses on producing accurate segmentation of a set of images at different scales. In the process of image co-segmentation, we turn our attention to the task of computing dense correspondences between a set of images. These correspondences are calculated in a dense grid of pixels, where each pixel is represented by an invariant descriptor computed at a unique, manually selected scale, this scale selection limits the efficiency of image co-segmentation methods when the common foregrounds appear at different scales. In this work, we use scale propagation to compute dense correspondences between images by assuming that if two images are being matched, scales should be assigned by considering feature point detections common to both images. We present both quantitative and qualitative tests, demonstrating significant improvements to segment images with large scale variation.