Large-Scale Coarse-to-Fine Object Retrieval Ontology and Deep Local Multitask Learning

Algorithm 3

Query expansion for image retrieval.

	Procedure: QUERY_EXPANSION
	Input:
	inputImgFeat: query image features
	startRetrievedImgFeatList: retrieved image feature list
	k: the number of retrieved images
	numOfExpansion: the number of query expansion applying times. Even though retrieval performance is better, it will have a trade-off with retrieval time. Recommended value: 3.
	Output:
	finalRetrievedImgFeatList: retrieved image feature list after moving over query expansion
	Goal:
	Enhance retrieval results
	BEGIN
	finalRetrievedImgFeatList = startRetrievedImgFeatList    #assignment retrieval feature list
	#do query expansion
	for i = 0 to numOfExpansion do:
	begin
	meanImgFeat = mean(finalRetrievedImgFeatList, inputImgFeat)  #compute mean vector from the retrieval feature list; this vector contains the most general features between all retrieval results and can be used to compute distance to all input features. The distance helps identify outliers and downrank them using an appropriate threshold.
	finalRetrievedImgFeatList = retrieve(meanImgFeat, k)  #compute the distance between mean vector and features of each retrieved image; reranking is applied based on the computed distances. Finally, take k first similar images as retrieval results.
	end
	END