A Multi-round Global Performance Evaluation Method for Interactive Image Retrieval

In interactive image retrieval systems, from the image search results, a user can select an image and click to view its similar or related images until he reaches the targets. Existing evaluation approaches for image retrieval methods only focus on local performance of single-round search results on some selected samples. We propose a novel approach to evaluate their performance in the scenario of interactive image retrieval. It provides a global evaluation considering multi-round user interactions and the whole image collection. We model the interactive image search behaviors as navigation on an information network constructed by the image collection by using images as graph nodes. We leverage the properties of this constructed image information network to propose our evaluation metrics. We use a public image dataset and three image retrieval methods to show the usage of our evaluation approach.