Extensive analysis of feature selection for compact descriptor

Feature selection for low bit rate mobile augmented reality applications

“…But the number of matched local feature pairs had been increased slightly which indicated that more true positive local features had been selected at the expense of computation. This observation is consistent with the result of [20,21]. However, extra computation is required by the combination methods as four selection metrics are needed to be computed for each feature, which will cause extra system delay when the feature number is large.…”

“…image pairwise matching accuracy) in the literature [15,20,21] while the position of the retrieved correct content in the retrieval list is not considered. However, the position of the correct content in the retrieved list is significant for MAR applications because MAR directly triggers the display of the first returned content.…”

“…The relevance of the output parameters of the SIFT detector, including the Differenceof-Gaussian (DOG) response θ peak (denoted as peak in the following paragraphs), scale , orientation θ orientation , location θ distance (the distance from the keypoint to the image center), are used for SIFT feature selection. Parameter θ peak is superior for identifying the most significant features for pairwise feature matching compared to other parameters [20,21].…”

“…But, instead of using the correctly matched features from a matching system directly, a posterior probability of a relevance parameter associated with correctly matched features is learnt offline from a training dataset. The training dataset contains both distorted images and corresponding clean images [20,21]. This posterior probability of a relevant parameter associated with correctly matched features is known as the relevance.…”

Quality of experience-based image feature selection for mobile augmented reality applications

“…Such differences are characterized during the feature detection and implied in the output of the feature detector. The output parameters including the Difference-of-Gaussian (DOG) response (denoted as peak in the following paragraphs), scale , orientation , location (the distance from the feature location to the image center) are evaluated individually to investigate the relevance score of these quantities to correctly matched pairs [12] as well as their combination [13] using the probability mass function of correctly matched features learned from dataset. Then, the features are selected on the basis of a relevance score.…”