Co-localization in noisy images through minimizing a ratio function

Object co-localization is a recently proposed vision problem to jointly localize the target object in a collection of images. The main practical challenge for co-localization is the existence of noisy images, in which the target object may be absent. Previous study relied on a prior estimate of the number of noisy images given by human. However, this prior knowledge would be somewhat too strong in practice. To improve on this, we propose a novel formulation for object co-localization in a noisy image collection, which does not rely on any prior knowledge of the noisy images. This is done by incorporating the object co-localization and noisy image identification jointly into a ratio-form objective. We develop an efficient algorithm based on Newton's method for ratio optimization, which can converge to the global optimum in only a few iterations. Extensive experiments conducted on two public benchmarks show that our approach can achieve better or comparable performance compared with several state-of-the-arts, and is robust to different proportions of the noisy images.