Citation Context

...ake into account the pairwise relationship of samples in the hash function learning. As a popular formulation, many recent works, including binary reconstructive embedding [11], complementary hashing =-=[23]-=-, supervised hashing with kernels [14], and iterative quantization [7] fall into the category of pairwise methods. Finally, semi-supervised hashing method plays a tradeoff between supervised informati...

Citation Context

...ecent hashing methods often employ machine learning techniques to leverage supervised information like pairwise constraints to design more efficient hash codes to search semantically similar entities =-=[35, 39, 22, 28]-=-. The key for the learning based hashing methods is to approximate the semantic similarity between entities in the learned Hamming space. However, an important but often neglected intuition is that si...

Citation Context

...though the latter does. 2.1. Tree-based approach FLANN is a representative tree-based method [20]. It automatically selects the best method among the randomized kd-tree [24], the hierarchical k-means [21] and the brute-force search, and tunes parameters for a given dataset. In the experiments, randomized kd-tree and hierarchical k-means were compared with the proposed method as FLANN itself did not work for large datasets. 2.2. Hash-based data independent approach Locality sensitive hashing (LSH) is a representative hash-based data independent method [4, 1]. There are many improvements (e.g., [19, 30]). However, its performance is known to be worse than the data dependent indexing methods. We also confirmed that LSH is far slower than the data dependent methods that follow. 2.3. Hash-based data dependent approach The data dependent methods can be categorized into those performing the selection of NN candidates in the Euclidean space and the Hamming space. Selection in the Euclidean space In this category, the methods use quantization and data compression techniques. Vector quantization (VQ) is used in some ANNS methods such as VQ-index [27] and IVFADC [12]. Product quantization (PQ) is use...

Citation Context

...LSH, the concept of “small code” or “compact code” has become the focus of many researchers from the CV community, and many learning-based hashing methods have come in to being [135][125][126][30][83]=-=[139]-=-[127][29][82][130][31]. These methods aim at learning the hash functions for better fitting the data distribution and labeling information, and thus overcoming the drawback of LSH. This part of the re...

Citation Context

...etrieval using functions - has become attractive tosresearchers as an alternative.sThere have been differentsapproaches as to which hashing function best minimizes thesdistance between neighbor pairs =-=[3, 12, 13, 14]-=-, but it seemssspectral hashing performs well. Spectral Hashing takes asmore holistic approach constructing hash functions based onsboth the structure of the data and ideal hash properties.sForsa form...

Citation Context

...turn sufficient reliable images. For performing an effective and efficient visual similarity search, we apply ap121 proximate nearest neighbor search approach using multiple complementary hash tables =-=[125]-=-. Besides, we also do an ANN search over the reference database and discover 10 most similar images for each reference image, which is regarded as the visual expansion of the reference image. The sear...

Citation Context

... save some computational costs, but can not solve the problem fundamentally. Hashing has been successfully applied to several nonNLP problems including object recognition [9][10], image retrieval [11]=-=[12]-=- and image matching [13][14]; however, it received limited attention in NLP fields. The general idea of hashing is to represent each document as a binary code (1-bit of a binary code is one digit of “...