Semi-supervised learning based on group sparse for relative attributes

Attribute-correlated local regions for deep relative attributes learning

“…Most of the previous works relevant to relative attributes depend on handcrafted features. 7,27,28,40 Recently, deep neural networks have also been extended for ranking applications. 35,36,41 Yaser et al 35 introduced a CNN-based model, which is composed of a feature learning and extraction part and a ranking part, to predict relative attributes.…”

“…Moreover, we cannot ensure that all the classified image sets contain only the given comparative image pairs, and so far, we have not found a dataset that satisfies this condition. Therefore, we start by sorting the images of S p in a descending order, using a group sparse-based semisupervised learning approach proposed by Hongxue et al 28 The ranked image collection is S 0 p ¼ fI 0 1 ; I 0 2 ; : : : ; I 0 m g. To initialize a single chain, we take the top N init images and select one patch from each image to form a patch set P ¼ fP 1 ; P 2 ; : : : ; P N init g. The appearance of each patch varies smoothly with its neighbors in the chain by minimizing the following objective function: E Q -T A R G E T ; t e m p : i n t r a l i n k -; e 0 0 6 ; 3 2 6 ; 3 7 3…”

“…Our experiments are evaluated on three public datasets: LFW-10, 30 Zappos50K-1, 29 and Shoes. 32 LFW-10 30 is a subset of the Labeled faces in the wild (LFW) dataset, which has 2000 images (1000 for training and 1000 for testing) and 10 attribute annotations, with 500 pairs of training and testing images per attribute.…”

“…We use the same training/testing split as that in Ref. 29. Shoes 32 dataset contains 14658 shoe images and 10 attributes, of which three are overlapping with Zappos50K-1: "open," "sporty," and "pointy."…”

“…As opposed to predicting the presence of an attribute, a relative attribute indicates the strength of an attribute in an image, and the relative descriptions are more precise and informative than the binary ones. Some representative relative attributesbased works have been proposed, from which Parikh and Grauman 27 designed more complex and task-specific models on their seminal work; however, the handcrafted visual features [28][29][30][31][32][33][34] are employed. Recently, the deep feature representations learned from CNN-based models have been exploited to predict relative attributes.…”