Saliency-Guided Multiple Class Learning

Overview of the Problem

Figure 1: An overview of the unsupervised object class discovery problem. The input is the same for different types of algorithms as a set of unlabeled images; the outputs of the algorithms may vary due to their different purposes.


Figure 2: Clustering and localization results from SIVAL dataset.



We tackle the problem of common object (multiple classes) discovery from a set of input images; we assume the presence of one object class in each image; this problem is loosely speaking unsupervised since we do not know a priori about the object type, location, and scale in each image. We observe that the general task of object class discovery in a fully unsupervised manner is intrinsically ambiguous; here we adopt saliency detection to propose candidate image windows/patches to turn a problem of unsupervised learning into a weakly-supervised learning task.

We propose an algorithm for simultaneously localizing objects and discovering object classes via bottom-up (saliency-guided) multiple class learning (bMCL). Our contributions are three-fold: (1) we adopt saliency detection to convert unsupervised learning into multiple instance learning, formulated as bottom-up multiple class learning (bMCL); (2) we propose an integrated framework which performs localizing objects, discovering object classes, and training object detectors simultaneously; (3) we demonstrate significant improvements of our framework over existing learning-based methods for multi-class object discovery and evident advantages of our formulation over competing methods in computer vision. In addition, although saliency detection has recently attracted many attentions, its practical usage for high-level vision tasks is yet to be justified. Our method validates the usefulness of saliency detection in dealing with "noisy inputs" for a top-down method to extract the common patterns.





The conference version of our paper on saliency-guided multiple class learning is available here. The supplementary material is available here.

The datasets we used include: SIVAL, iCoseg, and 3D Object Category. The random partition we used for experiments on the SIVAL dataset is available here.

The source code will be released soon.