Co-localization with Category-Consistent CNN Features and Geodesic Distance Co-Propagation

Co-localization is the problem of localizing categorical objects using only positive sets of example images, without any form of further supervision. This is a challenging task as there is no pixel-level annotations. Motivated by human visual learning, we find the common features of an object category from convolutional kernels of a pretrained convolutional neural network (CNN). We call these category-consistent CNN features. Then, we co-propagate their activated spatial regions using superpixel geodesic distances for localization. In our first set of experiments, we show that the proposed method achieves state-of-the-art performance on three related benchmarks: PASCAL 2007, PASCAL-2012, and the Object Discovery dataset. We also show that our method is able to detect and localize truly unseen categories, using six held-out ImagNet subset of categories with state-of-the-art accuracies. Our intuitive approach achieves this success without any region proposals or object detectors, and can be based on a CNN that was pre-trained purely on image classification tasks without further fine-tuning.