Citation Context

...ree level, so that the darkest values (ratio≈1) fall near the boundaries between k-means regions. cal k-means tree similar to that proposed by Fukunaga and Narendra (Fukunaga and Narendra, 1975). In (=-=Leibe et al., 2006-=-) the authors propose an efficient method for clustering and matching features in large datasets. They compare several clustering methods: k-means clustering, agglomerative clustering, and a combined ...

Citation Context

...e – there do not exist separate clusters. Existing solutions which improve over standard k-means include the use of mean-shift based clustering [9], hierarchical kmeans [21], agglomerative clustering =-=[14]-=-, and randomized trees [20]. They all present a compromise between accuracy and speed. In this paper we introduce an accurate and fast solution which differs significantly from existing solutions, as ...

Citation Context

...sely in linear time. Related work. Lazebnik et al. [8]alsoperformdiscriminativelearningtooptimize k-means, but are limited to small dictionaries and visual words which are Voronoi cells. Leibe et al. =-=[9]-=-alsoperformcompression,butnotinadiscriminative sense. Finally, Winn et al. [10]dodiscriminativecompressioninasimilar fashion, but we show that we perform better and can scale to larger dictionaries. F...

Citation Context

...ing. The SIFT descriptors are extracted from the training data as described in sections 2 and 3. The vocabulary trees are then constructed with the agglomerative clustering similar to the one used in =-=[9]-=-. Initially each descriptor forms a cluster. Two nearest clusters in the whole set are merged at each iteration based on their Euclidean distance. We continue merging until the number of clusters is 1...

Citation Context

...is to describe regions with a compact descriptor such as SIFT [9] or some discretization of it (e.g. ”visual words” [17]) and to store database image representations in a search tree (k-d [9], metric =-=[7]-=-, k-means [13, 16, 3]). 1 1 Terms ”viewpoint invariant features”, ”interest points”, ”patches”, ”distinguished regions” also appear in the literature. A better estimate of correspondence quality (a pr...

Citation Context

...ormation to estimate the covariance. It is a strong assumption but it may lead to better results in particular when used in combination with the simulated data. We employ the agglomerative clustering =-=[6]-=- to obtain the feature clusters. It is a bottom up clustering approach which iteratively finds the nearest features in the set and merges them into clusters. In our implementation features whose dista...

Citation Context

... of large scale recognition methods [27]. The possibility to match SIFT-like descriptors by sublinear tree-based search methods is one of its key properties and has been explored in different systems =-=[18, 21, 23, 27]-=-. We therefore carefully compare SIFT and M-SIFT performance with different search structures, variants of kd-trees and metric (ball) trees. Besides demonstrating that M-SIFT outperforms SIFT, valuabl...

Citation Context

...hical clustering (divisive clustering [9] as well as agglomerative clustering), clustering based on function optimization (e.g. expectation maximization (EM) type clustering [15]) or mixed techniques =-=[8]-=-. A general overview about clustering algorithms can be found in [15]. The two most commonly used techniques for codebook generation are agglomerative hierarchical clustering as well as K-means cluste...

Citation Context

...s k-means focus on clustering the most frequent descriptors, that are in general less discriminative. An algorithm similar to mean-shift was proposed to overcome this limitation. For similar reasons, =-=[12]-=- suggests to use agglomerative clustering which gives more compact clusters, while being more robust to outliers. Hierarchical clustering is also very efficient in terms of the computational cost to a...