Locality Preserving Hashing

Authors: Kang Zhao, Hongtao Lu, Jincheng Mei

AAAI 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental comparisons with other state-of-the-art methods on two large scale datasets demonstrate the effectiveness and efficiency of our method.
Researcher Affiliation Academia Kang Zhao, Hongtao Lu and Jincheng Mei Key Laboratory of Shanghai Education Commission for Intelligent Interaction and Cognitive Engineering Department of Computer Science and Engineering, Shanghai Jiao Tong University, China {sjtuzk, htlu, jcmei}@sjtu.edu.cn
Pseudocode Yes Algorithm 1 Locality Preserving Hashing (LPH) Input: Data X (zero-centered); random initialization matrix W(0); positive parameter ρ; number of hash bits K; iteration counter t 1. repeat Update binary codes Y(t) from W(t 1) by Eq.(12); Update projection matrix W(t) from Y(t) by Eq.(15); t t + 1; until convergence; Output: Hash codes Y and projection matrix W.
Open Source Code No The paper does not contain any explicit statement or link indicating that the source code for the described methodology is available.
Open Datasets Yes We evaluate our LPH method on the two benchmark datasets: STL-10 and ANN-GIST1M. STL-10 dataset is an image recognition dataset with higher resolution (96 96). It has 10 classes, 5000 training images, 8000 test images and 100000 unlabeled images. In our experiments, images are represented as 384-dimensional grayscale GIST descriptors (Oliva and Torralba 2001). The training set consists of 100000 unlabeled images, and the test set consists of 1000 test images. ANN-GIST1M dataset is a subset of the largest set provided for the task of ANN search. We randomly sample 100000 images from its 1 million 960-dimensional GIST features as the training set, and 1000 query images as the test set.
Dataset Splits No The paper mentions training and test sets but does not explicitly describe a validation set or provide details about how data was split for validation, such as percentages or sample counts.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory specifications) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers.
Experiment Setup Yes In practice, we use 50 iterations for all experiments, which has already achieved good performance. ... if we set ρ in Eq.(8), our method is equivalent to only minimizing the quantization loss (7), which is denoted as LPH-q in the experiments. These two paradigms can be seen as the special cases of our model. We experimentally present that our method (with ρ equal to 1) has the best performance. ... We begin the iterative algorithm with a random orthogonal initialization of W, which provides better performance than arbitrary random initialization since it corresponds to the orthogonality constraints. ... The gradient under canonical metric is given below: ... And we accelerate it by Barzilai-Borwein (BB) step size as in (Wen and Yin 2013).