On the burstiness of visual elements

Abstract

Figure 1. Illustration of burstiness. Features assigned to the most "bursty" visual word of each image are displayed. AbstractBurstiness, a phenomenon initially observed in text retrieval, is the property that a given visual element appears more times in an image than a statistically independent model would predict. In the context of image search, burstiness corrupts the visual similarity measure, i.e., the scores used to rank the images. In this paper, we propose a strategy to handle visual bursts for bag-of-features based image search systems. Experimental results on three reference datasets show that our method significantly and consistently outperforms the state of the art.

Rotation and translation covariant match kernels for image retrieval

“…γ n cos (n θ) (11) with γ 0 = (I 0 (κ) − e −κ ) 2 sinh (κ) and γ n = I n (κ) sinh (κ) if n > 0. (12) We design the feature map α(θ ), mappping an angle θ to a vector, as follows:…”

“…The final image vector obtained by each method is power-law normalized [11,15,40]. This processing improves the performance by efficiently handling the burstiness phenomenon.…”

“…This is critical when considering large datasets, as re-ranking methods depend on the quality of the initial short-list, which typically consists of a few hundred images. The initial stage is improved by better matching rules, for instance with Hamming embedding [9], by learning a fine vocabulary [10], or weighting the distances [11,12]. Moreover, it is useful to employ some geometrical information associated with the region of interest [9].…”

Improving the search accuracy of the VLAD through weighted aggregation of local descriptors

“…To remove effects of visual bursts [22], intra-normalization is proposed [12] where L 2 -normalization is independently applied to each subvector v i . It decreases the values of some bursty VLAD components, which can adversely dominate the similarity computation between VLADs.…”

Local consistent hierarchical Hough Match for image re-ranking

“…With this approach, the overhead of vocabulary can decrease greatly, however, it mainly focus on approximate near neighbour search. In addition, [7] surveys a case called burtiness that image has large number of repeated patterns are quantized to little words in images and solves it via a weight in similarity computing. [18] proposes a hypergraph-based approach to simultaneously utilize visual information and tags for image relevance learning.…”