Oracle in Image Search: A Content-Based Approach to Performance Prediction

  • Authors:
  • Liqiang Nie;Meng Wang;Zheng-Jun Zha;Tat-Seng Chua

  • Affiliations:
  • National University of Singapore;Hefei University of Technology;National University of Singapore;National University of Singapore

  • Venue:
  • ACM Transactions on Information Systems (TOIS)
  • Year:
  • 2012

Quantified Score

Hi-index 0.00

Visualization

Abstract

This article studies a novel problem in image search. Given a text query and the image ranking list returned by an image search system, we propose an approach to automatically predict the search performance. We demonstrate that, in order to estimate the mathematical expectations of Average Precision (AP) and Normalized Discounted Cumulative Gain (NDCG), we only need to predict the relevance probability of each image. We accomplish the task with a query-adaptive graph-based learning based on the images’ ranking order and visual content. We validate our approach with a large-scale dataset that contains the image search results of 1,165 queries from 4 popular image search engines. Empirical studies demonstrate that our approach is able to generate predictions that are highly correlated with the real search performance. Based on the proposed image search performance prediction scheme, we introduce three applications: image metasearch, multilingual image search, and Boolean image search. Comprehensive experiments are conducted to validate our approach.