Circular Reranking for Visual Search

0
910
Circular Reranking for Visual Search

Circular Reranking for Visual Search

Abstract of MAT LAB Project On Reranking For Visual Search System

MAT LAB Project On Reranking For Visual Search System.The rapid development of Web 2.0 technologies has led to the surge of research activities in visual search.
A common practice to improve search performance is to rerank the visual documents returned from a search engine using a larger and richer set of features.
The ultimate goal is to seek consensus from various features for reordering the documents and boosting the retrieval precision.
There are two general approaches along this direction:visual pattern mining and multi-modality fusion The former mines the recurrent patterns, either explicitly or implicitly, from initial search results and then moves up the ranks of visually similar documents.
Random walk for instance, performs self-reranking through identifying documents with similar patterns based on inter-image similarity and initial rank scores. This category of approaches, nevertheless, seldom explores the joint utilization of multiple modalities.

Conclusion

We have presented circular reranking which explores information exchange and reinforcement for visual search reranking.

Particularly, we analyze the placement of modalities in the circular framework which could lead to the highest possible retrieval gain in theory for search reranking.

 To verify our claim, we have presented approaches based on the existing works in the literature for predicting the modality importance to sort and weight the modalities accordingly for circular reranking.

Experiments conducted for image and video retrieval basically validate our proposal and analysis.

Performance improvement is also observed when comparing to other reranking techniques such as linear fusion based on oracle setting and fixed weights learnt from training examples.