Fine-Grained Visual Comparison Based on Relative Attribute Quadratic Discriminant Analysis
提出一种结合HOG和Gist特征的描述子，以及相对属性二次判别分析方法，用于降维和度量学习，在三个数据集上验证了细粒度视觉比较的效果。
In the vision research field, it is useful to make a comparison to see which one of two images exhibits a particular visual attribute more than the other. Feature representation and metric learning are two critical factors for a fine-grained visual comparison. In this paper, first an informative feature descriptor combining histograms of oriented gradient feature with gist feature is proposed, and then a method called relative attribute quadratic discriminant analysis is proposed for dimensionality reduction and metric learning simultaneously. In the method, those pairs most analogous to a test pair are identified by using the learned metric and then used to build a local ranking function. Finally, the test pair label is predicted by the ranking function for each attribute. Experimental results on the three databases of UT-Zap50K, Outdoor Scene Recognition dataset, and Public Figures faces dataset demonstrate the advantages of the proposed method.