PROJECT TITLE :

Discriminative Multi-View Privileged Information Learning for Image Re-Ranking

ABSTRACT:

Multi-view re-ranking algorithms often execute an asymmetrical comparison between the query image and the entire target image when calculating similarity for the query image. An inconsistency in the visual appearance can lead to a decrease in the retrieval accuracy, especially if a smaller portion of the image's ROI is used to represent the image objectness. Privileged Information (PI), a type of image prior that may accurately characterise the objectness of a photo, will be used to improve the accuracy of the multiview ranking system in this study. Discriminative multi-view ranking is proposed in order to achieve this goal, in which a unified training framework incorporates both the original global image visual contents and the local additional features to generate the latent subspaces with suitable discriminating power. Since multi-view PI features are not accessible for on-the-fly re-ranking, we merely project the original multi-view image representations into the latent subspace, and hence the re-ranking may be done by computing and sorting the distances to the separating hyperplane. Oxford5k and Paris6k public benchmarks show that our method gives an additional performance boost for accurate image re-ranking; a comparative research shows that our method is superior to existing multi-view methods.